Category id 582bffde79a0295e (Scout category id). 25 alternatives, shown in Scout's order.
Scout's category view
Scout's verdict: adopt. LiteLLM routes simple prompts to local models (free), complex to Claude. Real cost tracking replaces $0.35/dispatch estimates. Brief sent to withheld.
Org setup at the time, as Scout recorded it: No routing. All prompts go to Claude via Max subscription at flat rate.
Scout's judgement, from alternatives.json (2026-11-18). It is not a gaents rating.
LiteLLM is an open-source LLM API gateway (proxy server) that provides a unified OpenAI-compatible API for 100+ LLM providers including Anthropic, OpenAI, Azure, Google, AWS Bedrock, Cohere, HuggingFace, and local models via withheld/vLLM. It acts as a drop-in proxy: point your application at LiteLLM instead of the provider directly, and it handles routing, load balancing, failover, cost tracking, rate limiting,... (Scout's description)
Self-hosted open source: 0 (MIT license). Infrastructure: needs ~300-400MB RAM on existing VPS, plus PostgreSQL for logging (already running). Redis for caching (already running). Enterprise license (if needed for SSO/RBAC): custom pricing, estimated ~200-500 USD/month based on industry comparisons. For our use case (small team, no SSO needed): total 2-year cost = 0 GBP. Note: LiteLLM does not charge for the proxy itself — you pay your LLM providers directly. Operational cost is mainly the ~4 hours of setup and ongoing configuration maintenance.
Adopt when GPU is approved. LiteLLM's full value (routing simple prompts to local models, complex ones to Claude) depends on having local inference capability. Without GPU, it is still useful as a unified API with cost tracking and failover, but the ROI is lower. Deploy as part of the GPU compute plan (Week 2 per the plan). Zero license cost, just operational setup.
Limits for us
300-400MB RAM is significant on our single VPS — need to account for this in resource planning alongside GlitchTip, Loki, and other new services. The 3-4 second import time is annoying but only affects cold starts (proxy runs persistently). Database log accumulation will need periodic cleanup or archival — at our current volume (~1K-5K requests/day), the 1M log bottleneck is ~200-1000 days away but should be planned for. No built-in smart routing by prompt complexity (NadirClaw's differentiator) — routing is by model name, not by task difficulty. Blocked on GPU approval — without local models via vLLM/withheld, LiteLLM is just a pass-through proxy to cloud APIs, which is less valuable than a router splitting traffic between local and cloud. Python-based, which means the proxy adds latency (~5-10ms per request) compared to native API calls.
Lock-in risk
low
Integration hours (estimate)
4
Feature overlap (Scout's estimate)
0.0
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
RECOMMENDED. Unified API, cost tracking, rate limiting, load balancing, guardrails. Self-hostable proxy. [2026-10-19T22:08Z cloud refresh: Q4-2026 review: licence field corrected. Repo is active (last commit 2026-10-19). Verdict unchanged.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE: open core, core MIT (permissive), enterprise/ directory under its own licence; self_hostable and api_or_webhooks from README; data_export left out (no export doc read).]
NadirClaw
source-availablenot runnable on gaents yet
NadirClaw is an internal LLM model routing and orchestration system that manages the selection and routing of requests across multiple language model providers and instances. The system intelligently directs requests to appropriate models based on task requirements, cost constraints, latency goals, and model capabilities.
The platform maintains fallback chains and routing policies, enabling graceful degradation... (Scout's description)
CONFIRMED 2026-10-19: self-hosted NadirClaw core free under PolyForm Noncommercial; hosted Nadir Route takes '25% of the first $2,000 saved in a month, and 10% of every dollar above that'; commercial licence terms for self-hosting not priced on the page. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Scout, gaents-proven-2026-11-18.md A22026-11-18Scout review date
Scout's judgement
Verdict
adopt-now
Lock-in risk
low
Integration hours (estimate)
0
Feature overlap (Scout's estimate)
0.0
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Internal model routing system. Snake_case module. last_commit_iso nulled 2026-07-24 — was a Dec-2024 batch-placeholder artifact, not a real measurement (NULL beats a fabricated date; refresh live when the entry is next touched). | 2026-10-02 hygiene (Scout; gh api checks): public source-available repo found (655★, pushed 2026-10-01). No NadirClaw binary/package/service found on bigbrain; the 'internal, already deployed' claim appears only in our own files: UNVERIFIED, confirm with Fathom/Adam before keeping adopt-now. [2026-10-19T22:08Z cloud refresh: Q4-2026 review: licence record corrected. DB said 'Internal'. The repo LICENSE says it is PolyForm Noncommercial 1.0.0 and that commercial use 'requires a separate commercial license'. Quoted as fact; verdict left for Adam (house rule 3). Last commit 2026-10-19. Type left unchanged: the licence is source-available, which is not one of Scout's type values (commercial | open-source | freemium). ❓ OPEN, routed to Scout: add a type value or a licence-class field.] [2026-10-19T23:00Z cloud refresh: [2026-10-19 licence check] Repo LICENSE (HEAD 2026-10-19) is PolyForm Noncommercial 1.0.0. Text: 'COMMERCIAL use — including use in or by a for-profit business, or to build or operate a paid product or service — requires a separate commercial license.' type set to source-available under Scout ruling 1 (was the non-standard value 'proprietary'). self_hostable from README: 'source-available, self-hosted core'. CONFIRMED.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from README (OpenAI-compatible proxy); data_export from README (local JSONL + SQLite storage). licence_class and self_hostable already set.]
Olla
open-sourcenot runnable on gaents yet
Olla (github.com/thushan/olla) is a proxy and load balancer written in Go for self-hosted LLM servers. It routes OpenAI-compatible requests across nodes running withheld, LM Studio, llama.cpp, vLLM, SGLang, LiteLLM, LM Deploy, Docker Model Runner and others, with priority-based routing, automatic failover, KV-cache-aware sticky sessions, model discovery and a unified model catalogue. Two proxy engines: Sherpa... (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Local LLM inference, runs models on-device. last_commit_iso nulled 2026-07-24 — was a Dec-2024 batch-placeholder artifact, not a real measurement (NULL beats a fabricated date; refresh live when the entry is next touched). [2026-10-19T23:00Z cloud refresh: Naming fix (STATUS left-to-do 7): this row described and linked withheld (withheld.ai). Olla is a separate project by thushan. url changed to the repo; licence corrected from MIT to Apache-2.0 (LICENSE file at HEAD). Facts: last commit on default branch 2026-09-22; latest tag v0.0.29 (2026-08-10). The README says it 'works alongside API gateways like LiteLLM' and links a commercial owner page (tensorfoundry.io/products/olla). Pricing 'Free' and type open-source still correct. Verdict evaluate not changed (no evidence either way); its overlap with LiteLLM in this category is an open question for Adam.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from README (HTTP proxy for OpenAI/Anthropic-style requests); data_export n/a (stateless proxy).]
Portkey
open-sourcenot runnable on gaents yet
Portkey is a control panel for AI applications that builds on the OpenAI SDK, helping developers build reliable, secure, and production-ready AI apps easily. The Portkey Python SDK provides an AI Gateway layer enabling developers to write once and run with any provider, switching between any model from any provider seamlessly.
The SDK features automated fallbacks and retries across multiple models, load... (Scout's description)
CONFIRMED 2026-10-19: Open-source gateway free. Hosted: Developer 'Free Forever' (10k recorded logs/month, 3-day log retention); Production $49/month (100k recorded logs/month, +$9 per additional 100k requests up to 3M, 30-day log retention); Enterprise custom (10M+ logs/month). as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
GitHub stars
5200 as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
OpenAI SDK-compatible AI gateway with fallbacks and caching. last_commit_iso nulled 2026-07-24 — was a Dec-2024 batch-placeholder artifact, not a real measurement (NULL beats a fabricated date; refresh live when the entry is next touched). [2026-10-19T22:08Z cloud refresh: Licence corrected: the gateway repo LICENSE reads 'MIT License, Copyright (c) 2024 Portkey, Inc', not Apache-2.0 as stored (last commit 2026-05-25). Pricing now has concrete hosted tiers instead of 'Managed service available'. two_year_cost_gbp: agent-chose basis = hosted Production plan $49 x 24 months = $1,176 at £0.7502 per $1 = £882.24, no overages; self-hosting the OSS gateway is £0 software cost.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Portkey-AI/gateway LICENSE (MIT); open core: hosted platform is proprietary. self_hostable from gateway deployment docs; api_or_webhooks from README; data_export left out (hosted log export not checked).]
OpenAI (ChatGPT / GPT API)
commercialnot runnable on gaents yet
OpenAI sells ChatGPT seats and the GPT API. ChatGPT (OpenAI's own docs, 2026-10-19): Free; Go $8/month; Plus $20/month; Pro at $100, $200 or $500/month; Business $20/user/month billed annually ($25 monthly, 2+ users). API flagship prices per 1M tokens input / cached / output: gpt-6-astra $10 / $1 / $50; gpt-6.1-sol $2 / $0.10 / $10; gpt-6-luna $0.10 / $0.01 / $0.50. GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini... (Scout's description)
CONFIRMED 2026-10-19 (API): Flagship, Standard, short context, per 1M tokens input / cached / output: gpt-6-astra $10.00 / $1.00 / $50.00; gpt-6.1-sol $2.00 / $0.10 / $10.00; gpt-6-luna $0.10 / $0.01 / $0.50. GPT-4o no longer on the API pricing page. CONFIRMED 2026-10-19 (ChatGPT seats, OpenAI's own docs at learn.chatgpt.com/docs/pricing): Free $0; Go $8/month; Plus $20/month; Pro 'Plans at $100, $200, or $500 USD per month'; Business $20/user/month '2+ users, billed annually. $25 per user per month when billed monthly'; Enterprise & Edu 'Contact sales' (no list price). UK App Store in-app prices (iOS only): Plus £19.99, Go £6.99, Pro 100 £88.90, Pro 200 £200.00, Pro 500 £444.99. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Evaluate as a supplementary API layer routed via LiteLLM. GPT's strengths (1.5% hallucination rate, fast structured outputs) complement Claude's strengths (200K context, sustained reasoning). The pragmatic approach is 'both, not either-or' — route specific task types to the optimal model. Do not adopt ChatGPT Business as a product; adopt the GPT API as an inference option alongside Claude. Requires LiteLLM deployment first.
Limits for us
Our agent fleet runs on Claude Code — switching to GPT would require rewriting every agent's CLAUDE.md, tool integration, and operational patterns. The 150+ user Enterprise minimum locks us out of the most secure tier. ChatGPT Business at $25/user/mo is priced per human seat, not per agent — scaling to 85+ agents would cost $25,500/year minimum. The API is the viable integration path, but Claude's 200K context window and superior sustained reasoning make it the better choice for our agents' core tasks (build-vs-buy analysis, code review, strategic recommendations). GPT's advantage — lower hallucination rate and faster structured outputs — is real but narrow. The Assistants API deprecation in August 2026 creates migration risk for any workflow built on it. Hidden reasoning token costs on o1/o3 make budget planning unreliable. Best used as a supplementary layer via LiteLLM: route extraction and structured output tasks to GPT, keep reasoning on Claude.
Lock-in risk
medium
Integration hours (estimate)
8
Feature overlap (Scout's estimate)
0.5
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
From company-brain: wins on hallucination rate (1.5% vs Claude 4-10%), structured outputs (23-28% faster, 41-43% cheaper). Claude wins on context (200K vs 128K), sustained reasoning. 150+ user Enterprise minimum. Assistants API deprecated Aug 2026. Hidden reasoning token costs inflate 2-4x. Best as supplementary layer: GPT for extraction, Claude for reasoning. 2yr cost: GBP 4.3-215K. [2026-10-19T22:08Z cloud refresh: 2026-10-19 freshness check: GPT-4o deprecated. OpenAI deprecations page, 2026-10-14: 'we notified developers using certain GPT-4 series models of their deprecation and removal from the API on January 14, 2027.' GPT-4o (gpt-4o-2024-08-06, gpt-4o-2024-11-20) and GPT-4o mini shut down January 14, 2027; recommended replacement gpt-6-luna. Also: 'Priority processing was renamed Fast mode on July 30, 2026.' ChatGPT seat prices could not be re-read.] [2026-10-19T23:00Z cloud refresh: GPT-4o check: this entry is about OpenAI's whole API and ChatGPT, not GPT-4o alone, so the verdict is not changed. What changes: GPT-4o (gpt-4o-2024-08-06, gpt-4o-2024-11-20), GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down on the API on 2027-01-14; OpenAI's named replacement for all four is gpt-6-luna ($0.10 in / $0.50 out per 1M, Standard short context). gpt-4o-2024-05-13 shuts down earlier, on 2026-10-23 (replacement gpt-5.6-sol). The what_it_does text still quotes 'GPT-4o ($2.50/$10 per 1M tokens input/output)' and '128K' context: those are out of date. ❓ OPEN: rewrite what_it_does and re-base two_year_cost_gbp on the gpt-6 family.] [2026-10-19T23:00Z cloud refresh: 2026-10-19 session 2: openai.com/chatgpt/pricing and chatgpt.com/pricing still return HTTP 403 with four different user agents. Seat prices came from OpenAI's own product docs (learn.chatgpt.com/docs/pricing.md, linked from developers.openai.com/llms.txt) and the official UK App Store listing (apps.apple.com/gb/app/chatgpt/id6448311069, seller OpenAI OpCo, LLC). Pro is now three tiers ($100/$200/$500), not one $200 tier. The App Store also lists a 'ChatGPT Plus £200.00' item with no period stated; not interpreted. Enterprise has no published price.] [2026-10-20T01:32Z cloud refresh: what_it_does rewritten from the 2026-10-19 cloud facts (patches 05 and 07); deprecation dates re-checked 2026-10-20 on developers.openai.com/api/docs/deprecations. Removed the old GPT-4o $2.50/$10 price and the unsourced benchmark comparisons.] [2026-10-20T01:33Z cloud refresh: type paid -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class proprietary (closed hosted service under the OpenAI Services Agreement); api_or_webhooks from API reference; data_export from ChatGPT export help page; self_hostable left out for the service (OpenAI also publishes some open-weight models, so 'false' not stated).]
Grok (xAI)
commercialnot runnable on gaents yet
Grok is xAI's (Elon Musk's AI company) large language model, available as a chatbot on X (Twitter) and via API. Grok 4.1 offers a 2M token context window (the largest commercially available), strong math/coding benchmarks (93.3% AIME), and aggressive API pricing at $0.20/$0.50 per 1M tokens (input/output) for Grok 4.1 Fast — approximately 25x cheaper than Claude Opus. Business plan is $30/seat/month. SOC 2... (Scout's description)
CONFIRMED 2026-10-19: API cheapest text model still grok-build-0.1 at $1.00 in / $2.00 out per 1M tokens (short context); flagship grok-4.7 $2.00 / $6.00. Consumer/team plans are now 'The X Subscription': Free $0; Plus $8/mo; Premium $30/mo; Super $100/mo; Ultra $200/mo. Teams: Premium $30/seat, Super $100/seat, Ultra $200/seat. Existing SuperGrok plans continue. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Skip. Grok's technical capabilities are real but irrelevant given the catastrophic safety failures and reputational risk. A healthcare PE fund cannot use an AI tool documented by RAND, CNN, and multiple governments as having the worst safety record in the industry. The 16,000-32,000 GBP cost is secondary to the existential reputational risk. Revisit only if xAI demonstrates sustained safety improvement AND obtains HIPAA compliance, neither of which appears likely in 2026.
Limits for us
The safety crisis is disqualifying for a healthcare PE fund. Generating content attributed to or associated with Grok creates reputational risk with LPs, regulators, and portfolio companies operating in regulated healthcare markets. No HIPAA compliance eliminates use for any healthcare-adjacent work. Beyond safety: Grok's advantages (2M context, cheap API) do not outweigh the risks — Claude's 200K context is sufficient for our use cases, and Groq offers cheaper inference without the reputational baggage. The Elon Musk association alone makes Grok-branded outputs unsuitable for LP communications, investor updates, or any external-facing materials. Even internal use risks normalising a tool with documented child safety failures. The only potentially defensible use case — real-time X monitoring via API — can be achieved through other means (Brave Search API, direct X API).
Lock-in risk
medium
Integration hours (estimate)
4
Feature overlap (Scout's estimate)
0.3
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
From company-brain: technically impressive but poor strategic fit for healthcare PE. Safety failures ('among worst AI chatbots assessed' - Common Sense Media). No HIPAA. Elon Musk associations risky for LP-facing materials. Potential supplementary: Scout for real-time X monitoring. Revisit in 12-18 months. 2yr cost: GBP 16-32K. · 2026-09-14 PRICE CORRECTION (S38): this entry read 'Grok 4.1 Fast $0.20/$0.50' — stale by a model generation and ~6x LOW. STRUCTURAL FINDING: the supergrok CLI is NOT an inference lane — it authenticates by browser OAuth against a SuperGrok/X-Premium+ subscription while the API authenticates by key; they are DISJOINT billing surfaces and a SuperGrok subscription includes ZERO API credits. So the CLI is a Claude-Code competitor, not something LiteLLM can route. The API IS routable: https://api.x.ai/v1, OpenAI-compatible, drop-in. TUNABLES: `max_output_tokens` DEFAULTS TO 128,000 — one runaway grok-4.6 generation is $0.77; `reasoning.effort` has five levels incl. `xhigh`; `store` retains conversations 30 days by a default we never chose. Input price doubles above 200k context, repricing the whole call. · 2026-09-14 STALE-TRIGGER FIX (S38, L060): trigger read 'None foreseeable — safety concerns prohibitive' while Adam was routing 8 Grok items in one batch. A 'None foreseeable' trigger on a tool the principal keeps asking about is exactly the drift L060 exists to catch. [2026-10-19T23:00Z cloud refresh: 2026-10-19: docs now branded 'SpaceXSI (xAI)'. SuperGrok is replaced for new customers by The X Subscription, which also bundles Cursor (acquired by SpaceX Aug 2026). Old 'SuperGrok price NOT VERIFIED' now resolved.] [2026-10-19T23:00Z cloud refresh: 2026-10-19 session 2: resolves the old 'SuperGrok consumer tier price NOT VERIFIED' gap with the official App Store listing (apps.apple.com/us/app/grok/id6670324846, seller X Corp.). App Store prices can differ from web prices. xAI docs: 'New self-serve customers choose from The X Subscription plans. Existing Cursor, SuperGrok, and X Premium plans continue.' [merge: `pricing` is carried by patch 06. Fact to add there or in notes: the official Grok App Store listing shows legacy SuperGrok in-app prices, US: Lite $10.00, SuperGrok $30.00, Plus $100.00, Heavy $300.00 (UK: £10/£30/£100/£300). These are for existing SuperGrok plans; new customers start on The X Subscription.]] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from Grok API docs (REST reference, Batch API, WebSocket mode). licence_class left out: x.ai terms returned HTTP 403. self_hostable left out (hosted API, but some older Grok weights were published, so false would mislead). data_export left out. Docs site now titled 'SpaceXSI Docs'.]
NVIDIA build.nvidia.com (free tier)
freemiumnot runnable on gaents yet
build.nvidia.com is NVIDIA's free-tier API endpoint for ~133 listed (not all callable) open and commercial models accessed via OpenAI-compatible REST. The free tier runs at ~4-5 requests/second sustained with burst capacity up to ~131 requests before the first 429. Daily cap is not publicly documented.
Characterization by withheld (2026-04-24) validated 10/10 benchmark performance on Llama-3.3-70B (34 tok/s,... (Scout's description)
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
ADOPT-NOW. Free inference for models LISTED (133 claimed; 55/82 listed returned 404 on 09-22) via OpenAI-compatible API. The '~$900/mo displacement at 60% adoption' figure was NEVER measured (withheld 09-23); measured adoption 0. Avoid: GLM-5.1 (2 tok/s, use Z.ai), withheld (chat-template bug), withheld (EU timeouts), Llama-3.1-405B (10 tok/s long prompts). [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from NVIDIA developer terms (proprietary); api_or_webhooks from API quickstart; self_hostable left out (this row is the hosted free tier; quickstart says 'The NIM is hosted by NVIDIA'); data_export n/a (compute API).] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): self_hostable true from NVIDIA Run NIM Anywhere docs: the same NIM models offered on build.nvidia.com can be downloaded as containers and self-hosted (the free hosted tier itself runs on NVIDIA DGX Cloud). data_export n/a (inference API).]
Ling-2.6-1T
open-sourcenot runnable on gaents yet
Ling-2.6-1T is a trillion-parameter mixture-of-experts (MoE) language model open-sourced by InclusionAI under the MIT license. The model uses BailingMoE V2.5 architecture combining hybrid attention (multi-latent attention + linear components) with sparse routing to achieve 1 trillion parameters while maintaining inference efficiency through selective expert activation.
The weights are published on HuggingFace as... (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Ling-2.6-1T is a verified, published trillion-parameter MoE model from InclusionAI (MIT license). Unlike DeepSeek-V4 (S17 comparison: hype, unverified), Ling weights are publicly downloadable, safetensor format validated, 160-shard model file index published. BailingMoE V2.5 architecture (mixture-of-experts hybrid attention). Real model with real weights, not vaporware. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: HF lastModified 2026-09-24; HF licence tag mit; gated=false.]
freellmapi (tashfeenahmed)
open-sourcenot runnable on gaents yet
An OpenAI-compatible proxy that stacks the free tiers of 28 upstream LLM providers behind one /v1 endpoint. Ships a React+Vite admin dashboard, stores provider keys AES-256-GCM encrypted in SQLite, and does smart routing with automatic failover (rate-limit/5xx/timeout on provider A cascades to B, up to 20 attempts) plus per-key usage tracking to stay under each free-tier cap. Speaks OpenAI Chat Completions, the... (Scout's description)
Free self-hosted. Optional 'Premium live catalog' $19/yr at freellmapi.co — so the tweet's '100% free' is not accurate. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
GitHub stars
52399 as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Self-hosted OSS: 0 GBP direct cost. Optional Premium live catalog: 19 USD/yr ≈ 15 GBP → 30 GBP over 2 years. Real cost is legal/compliance risk from stacking provider free tiers — plausibly unbounded if a provider decides to enforce ToS. For Aubrains commercial use we already have LiteLLM (adopt-now) + NadirClaw (internal) + NVIDIA free tier (adopt-now, single-provider = clean ToS) covering the routing use case.
SKIP for commercial/agent-fleet use. Functional overlap with our existing LiteLLM + NadirClaw + NVIDIA stack is high, and the ToS risk of stacking 18 provider free tiers is unacceptable for commercial use. Recommend Adam file for personal side projects only if interested. No follow-up research warranted unless the provider-stacking model gets a clean-ToS variant.
Limits for us
The blocking issue is not technical, it is contractual. The README carries a dedicated 'Terms of Service review' section and the repo's own GitHub description ends with the words 'Personal experimentation only.' Several upstream providers — NVIDIA NIM's eval-only ToS is the clearest — prohibit production traffic served off a stacked free tier. Routing commercial agent-fleet workloads through this would breach individual provider terms, and the failure mode is not a bill, it is providers banning the accounts. Technically it also overlaps ~85% with LiteLLM, which we already run at adopt-now, and it depends on 28 upstream free tiers whose quotas change weekly — that is a fragile base for anything with an SLA. The $19/yr premium catalog exists precisely because keeping up with that churn is hard.
Lock-in risk
medium
Integration hours (estimate)
4
Feature overlap (Scout's estimate)
0.85
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
CONFIRM-EXISTING — already in the DB under Model Routing / LLM Proxy at verdict=skip, deep-enriched 2026-07-05. Field drift since: github_stars 15,170 -> 16,981; provider count 18 -> 28; endpoint count 161 -> 339; monthly token claim ~1.7B -> ~4B; last_commit_iso 2026-07-05 -> 2026-07-20; v0.5.0 released 2026-07-17. Verdict, lock-in, integration hours and action_readiness UNCHANGED — the ToS problem scaled up with the provider count rather than away. Patch numerics only (L003). [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-10-20; licence unchanged (MIT).]
withheld
freemiumnot runnable on gaents yet
100+ models one-billing-surface aggregator including premium (Claude/GPT/Gemini) and OSS. Free-tier includes one 262K-context model (Ling 3.0 Flash Fin). Model-shopping and one-off niche models the right entry point. (Scout's description)
CONFIRMED 2026-10-19: Plans Free / Standard / Business / Enterprise. Inference billed at provider list price on every plan (no per-request markup); platform fee charged on credit purchases: 5.5% Standard, 8% Business, Enterprise by contract. Free plan: 25+ free models, 50 requests/day. BYOK: $25,000/month list-price inference with no fee, 5% after. Model prices (from /api/v1/models): inclusionai/ling-3.0-flash $0.021/$0.063 per Mtok (no ':free' Ling variant listed); tencent/hy-mt2-1.8b $0.044/$0.177 per Mtok. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Wire into LiteLLM as fallback route; do NOT route Claude/GPT/Gemini through it — direct is cheaper because we hold seats/accounts. [2026-10-19T22:08Z cloud refresh: 2026-10-19 price check: stored 'Passthrough margin on premium models' replaced; withheld states it does not mark up provider pricing and charges a fee on credit purchases. Stored 'Ling 3.0 Flash free' no longer true: ling-3.0-flash is now paid and is cheaper than Hy-MT2-1.8B, so Hy-MT2 is no longer the cheapest. two_year_cost_gbp unchanged (null): usage-based.] [2026-10-19T22:08Z cloud refresh: Q4-2026 review: withheld announced on 2026-08-19 that it is joining Stripe (acquisition). It says the product and integration do not change. Verdict unchanged; watch for pricing changes under Stripe. Pricing for this entry is carried by patches/02-price-refresh.json (fuller text, same facts); apply 02 before 03.] [2026-10-20T01:33Z cloud refresh: type closed-source -> freemium (proprietary; pricing has free tier + paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Terms (closed service); api_or_webhooks from docs index (API reference). self_hostable left out (no page says cloud-only). data_export left out: docs show only an aggregated usage export (CSV/PDF), not all data.] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): self_hostable false: the Enterprise page (the page that lists enterprise options) gives only the hosted endpoint baseURL https://withheld.ai/api/v1 and BYOK provider keys; neither it nor the pricing page offers a self-hosted or on-prem edition (agent-checked both pages). data_export n/a as in pass 1 (only usage export).]
Together.ai
commercialnot runnable on gaents yet
Cheapest useful OSS-model aggregator + fine-tuning API + dedicated H100/B200 GPUs. DeepSeek V4 Flash at $0.14/$0.28 with $0.03/M cached is the current price leader for reasoning-quality models. Fine-tuning $0.48/M for <16B models. (Scout's description)
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
The price leader for OSS models. Reserved throughput 83% cheaper than commercial for sustained traffic — do the math when a workload steadies. [2026-10-20T01:33Z cloud refresh: type closed-source -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Terms of Service; api_or_webhooks from terms and docs index; self_hostable left out (no official page found either way); data_export n/a (inference API).]
DeepInfra
commercialnot runnable on gaents yet
One-billing-surface aggregator with the DIFFERENTIATOR that they carry Claude Fable-5 and Opus-5 alongside OSS models. Consistent H100/A100 hardware. Flex tier 0.8x for non-production; Priority 1.5x for latency-sensitive. (Scout's description)
Claude Fable-5 $10 in; Claude Opus-5 $5 in; Kimi-K3 $2.85 in; Gemini 3.1 Pro $2 in; Claude Sonnet-4-6 $3 in. Flex tier 0.8x for non-prod. Priority 1.5x. All H100/A100 hardware. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Passthrough margin on Anthropic is 10-20% over Anthropic direct list. Justified only when the billing seam matters more than the margin. [2026-10-20T01:33Z cloud refresh: type closed-source -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from DeepInfra terms; self_hostable false: docs.deepinfra.com describes a hosted inference cloud (hosted API, no self-hosted edition); api_or_webhooks from docs; data_export n/a (pure compute API).]
Fireworks.ai
commercialnot runnable on gaents yet
Enterprise-oriented aggregator whose differentiator is the SERVERLESS TRAINING API (only aggregator that has this) and VLM fine-tuning with image token billing. $1 free credits to start; embeddings from $0.008/M. (Scout's description)
$1 free credits. Embeddings $0.008-$0.10 per M. Fine-tuning $0.50-$40/M (SFT/DPO/RL). H100 $8/hr, B200 $13/hr, B300/GB300 $15-20/hr. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
H100 at $8/hr is 2x Together's $3.99/hr — buy Fireworks only for their specific stack (fine-tune ecosystem), not for on-demand compute. [2026-10-20T01:33Z cloud refresh: type closed-source -> commercial (no recurring free tier: $1 free credits (one-off); rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from docs; licence_class left out (terms page would not load: download triggered); self_hostable left out; data_export n/a (inference API).] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): licence_class proprietary from the Terms of Service PDF (fireworks.ai/terms-of-service redirects to a PDF, Last Updated July 10, 2026; read with pdftotext). self_hostable true from the Enterprise page (bring your own cloud and hardware); this is an enterprise option, not the self-serve tier. data_export n/a (inference API).]
Google Gemini API (AI Studio)
freemiumnot runnable on gaents yet
Google's direct Gemini API via AI Studio. OpenAI-compatible surface available, so LiteLLM can route to it. Generous free tier that scales per key. (Scout's description)
CONFIRMED 2026-10-19: Gemini 2.5 Flash-Lite is no longer listed on the official pricing page. Cheapest Flash-Lite listed: Gemini 3.1 Flash-Lite $0.25/M input (text/image/video), $0.50/M audio, $1.50/M output, cache $0.025/M. Gemini 3.5 Flash-Lite $0.30/M in, $2.50/M out; Batch and Flex $0.15 / $1.25. Free tier: 'Free of charge' input and output on these models (rate limits not on this page). Grounding with Google Search: 5,000 free requests/month shared across Gemini 3+ models, then $14 per 1,000. The >200k-prompt price step is shown for Gemini 2.5 Pro ($1.25 to $2.50 input). as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Price leadership is established; QUALITY on our workloads is not — this is an `evaluate` on a price finding, not an adopt. Needs a bake-off before it becomes a routing default. Whether Gemini 2.5 Pro remains free-tier could not be settled: Google's own pricing page and third-party trackers disagree as of 2026-09-14.
Lock-in risk
low
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
TUNABLES (Adam's standing redirect — these are the defaults that cost money): BATCH ENDPOINT is a flat 50% discount and we use it NOWHERE — the single biggest free saving available on this provider; THINKING TOKENS bill at the OUTPUT rate and the budget is unbounded by default — against 3.5 Flash's $9.00/M out this is the most expensive default in the provider; cache TTL is a real cost knob at $0.50-1.00 per M-token-hour of storage; context >200k DOUBLES input price for the entire call. POSITIONING: build.nvidia.com stays the default FREE lane (the '~$900/mo displacement' figure was never measured, withheld 09-23); Gemini free (~360M tok/mo/key ~ $370/mo list) is the correct SECOND free lane, and unlike NVIDIA it scales per key. [2026-10-19T22:08Z cloud refresh: Changed: the stored lane (Gemini 2.5 Flash-Lite ~$0.127/M at 10:1 blend, cache $0.01/M) cannot be confirmed: 2.5 Flash-Lite no longer appears on the pricing page. The deprecations page still lists gemini-2.5-flash-lite (release July 22, 2025) with 'No shutdown date announced'. At the stored 10:1 input:output blend, 3.1 Flash-Lite works out to (10 x $0.25 + $1.50) / 11 = $0.364/M. That is arithmetic on list prices, not a measured cost. The free-tier figure of ~360M tokens/month/key is not stated on the page, so it is UNVERIFIED. Page banner: 'Gemini 4 Argon is now available.'] [2026-10-20T01:33Z cloud refresh: type commercial-saas -> freemium (proprietary; pricing has free tier + paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Gemini API additional terms (proprietary); api_or_webhooks from Gemini API docs; self_hostable left out (no page read says cloud-only); data_export n/a (compute API).]
harness-bridge
open-sourcenot runnable on gaents yet
See research/tweet-triage-2026-10-02-scout.md (section for harness-bridge). (Scout's description)
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/tweet-triage-2026-10-02-scout.md (section for harness-bridge).
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Merged from research/tweet-triage-2026-10-02-scout.md by merge_triage_table.py. [2026-10-20T01:33Z cloud refresh: type None -> open-source (license text: MIT; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-09-27; licence unchanged (MIT).]
claude-code-router
open-sourcenot runnable on gaents yet
See research/tweet-triage-2026-10-02-scout.md (section for claude-code-router). (Scout's description)
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/tweet-triage-2026-10-02-scout.md (section for claude-code-router).
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Merged from research/tweet-triage-2026-10-02-scout.md by merge_triage_table.py. [2026-10-20T01:33Z cloud refresh: type None -> open-source (license text: MIT; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable from install docs; api_or_webhooks = local HTTP gateway routes. data_export n/a (proxy, holds no user data).]
Cloudflare AI Gateway
freemiumnot runnable on gaents yet
See research/cloudflare-offerings-deep-review-2026-10-02.md (section for Cloudflare AI Gateway). (Scout's description)
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/cloudflare-offerings-deep-review-2026-10-02.md (section for Cloudflare AI Gateway).
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Merged from research/cloudflare-offerings-deep-review-2026-10-02.md by merge_triage_table.py. [2026-10-19T22:08Z cloud refresh: Changed: stored 'Logpush $0.05/M over 10M' is the Workers Trace Events Logpush rate. The AI Gateway pricing page (updated Oct 15, 2026) now says AI Gateway Logpush is billed by the GB-based Logpush pricing (page updated Oct 12, 2026). Also new on the page: DLP free, Guardrails billed as Workers AI.] [2026-10-20T01:33Z cloud refresh: type None -> freemium (proprietary; pricing has free tier + paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Cloudflare terms; api_or_webhooks from REST API docs; data_export from docs (gateway logs exportable via Logpush); self_hostable left out (no page found).] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): self_hostable false: Cloudflare Self-Serve Subscription Agreement says its Services are Cloudflare cloud-based solutions that require a Cloudflare account; product docs give only Cloudflare-hosted endpoints and list no self-hosted or on-prem edition (agent-checked docs overview).]
Cloudflare Workers AI
freemiumnot runnable on gaents yet
See research/cloudflare-offerings-deep-review-2026-10-02.md (section for Cloudflare Workers AI). (Scout's description)
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/cloudflare-offerings-deep-review-2026-10-02.md (section for Cloudflare Workers AI).
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Merged from research/cloudflare-offerings-deep-review-2026-10-02.md by merge_triage_table.py. [2026-10-20T01:33Z cloud refresh: type None -> freemium (recurring free allowance: 10K Neurons/day free; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Cloudflare terms (service is closed; many models are open weights); self_hostable false: Workers AI docs say models run on Cloudflare's network (hosted API); api_or_webhooks from REST API docs; data_export n/a (pure compute API).]
Z.ai API (GLM-5.3 / withheld)
commercialnot runnable on gaents yet
Pay-per-token API for Zhipu's GLM-5.3 family, with an Anthropic-compatible endpoint, so existing Claude/OpenAI clients can point at it with a new base URL and key. (Scout's description)
GLM-5.3 $1.40 in / $0.26 cached / $4.40 out per 1M; Flash $0.15 / $0.03 / $0.50; FlashX $0.37 / $0.075 / $1.25 (read 10-08) as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Data: z.ai DPA says prompts/outputs are not stored (no switch needed). The announced ZDR option is for bigmodel.cn (China), not z.ai. Contracting entity JINGSHENG HENGXING TECHNOLOGY PTE.LTD, Singapore law, SIAC arbitration. Subscription terms §4.3: plan licensed to ONE natural person, no sharing with 'any organization'. §4.2: no calling the models from your own bots, apps or systems; no reselling. API terms §1(f)(iii)-(iv): excludes finance, investment, legal services and decision-making. withheld plans to measure it on our tasks before any switch. [2026-10-20T01:33Z cloud refresh: type paid -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class proprietary for the hosted API (Z.ai Terms of Use; open GLM weights are separate); api_or_webhooks from quick start; self_hostable left out (open weights exist, row is the API); data_export n/a (inference API).]
GLM Coding Plan (individual)
commercialnot runnable on gaents yet
Flat-fee GLM access for coding tools, including Claude Code via env vars. (Scout's description)
CONFIRMED 2026-10-19: from 18 USD/month (Lite). Lite/Pro/Max quotas: 2,000/12,000/28,000 credits per 5 hours. UNVERIFIED: Pro $80 / Max $168 and quarterly/yearly discounts (z.ai/subscribe is JS-rendered; not in docs text). as of 2026-10-19unverified in Scout's data
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
SKIP for agents on the terms, not the tech: Subscription terms §4.3: plan licensed to ONE natural person, no sharing with 'any organization'. §4.2: no calling the models from your own bots, apps or systems; no reselling. API terms §1(f)(iii)-(iv): excludes finance, investment, legal services and decision-making. Individual plan data is kept while the account exists and may be used for training on non-personal content (ToU §IV.3(a)). Gotcha: the documented Claude Code setup routes everything to withheld unless you map model names. [2026-10-19T23:00Z cloud refresh: 2026-10-19 price refresh: Lite $18 entry price confirmed on docs.z.ai; higher tiers not readable.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Z.ai Terms of Use. api_or_webhooks: the plan is used through an Anthropic-compatible API endpoint (Claude Code setup doc). self_hostable left out (no page says cloud-only). data_export n/a (pure compute API).]
GLM Coding Team Plan
commercialnot runnable on gaents yet
Seat-based GLM coding subscription for organisations; the lawful flat-fee route if GLM wins the head-to-head. (Scout's description)
UNVERIFIED: Seat types are now 'Standard Seat' and 'Premium Seat' (credits-based plan since 2026-07-30), minimum 2 seats, no mixed seat types. Standard Seat: 15,000 credits per 5 hours, 66,000 per week; Premium Seat: 35,000 per 5 hours, 155,000 per week; off-peak usage at 50% credit rate. Per-seat prices are shown only on https://z.ai/subscribe, which renders with JavaScript and could not be read; the stored '$88 / $188 per seat' could not be confirmed. as of 2026-10-19unverified in Scout's data
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Check whether its terms also bar automated/agent use before buying (the §4.2 bot clause is in the subscription terms). No refunds. [2026-10-19T22:08Z cloud refresh: 2026-10-19 check: Z.AI relaunched GLM Coding Plan as a credits-based plan. Official Plan Update Announcement (published July 30, 2026): 'GLM Coding Plan is moving to a credits-based system ... The new credits-based plan is now available. Previous plans are no longer sold to new users.' Existing team subscriptions keep their terms until the end of the billing cycle. Stored seat names 'Pro'/'Max' replaced by 'Standard Seat'/'Premium Seat'. two_year_cost_gbp unchanged (null): current price not readable.] [2026-10-20T01:33Z cloud refresh: type paid -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Z.ai terms of use (proprietary); api_or_webhooks from Team Plan docs (plan API key); self_hostable left out; data_export n/a (compute subscription).]
Kimi membership (Kimi Code, K3)
commercialnot runnable on gaents yet
See research/tweet-triage-2026-10-14-scout.md (section for Kimi membership (Kimi Code, K3)). (Scout's description)
CONFIRMED 2026-10-20: international web pricing (kimi.com/membership/pricing, USD): Adagio free $0; Plus $19/month or $15/month billed annually ($180/yr); Pro $39 or $31 ($372/yr); Max $99 or $79 ($948/yr); Ultra $199 or $159 ($1,908/yr). Go is not offered on the international page (Go is a China plan: ¥49/month, no Kimi Code). China plans per help page: Go ¥49, Plus ¥99, Pro ¥199, Max ¥699, Ultra ¥1,399 a month (annual ¥39/¥79/¥159/¥559/¥1,119 a month). Legacy Andante/Moderato/Allegretto/Allegro plans continue for existing subscribers. Earlier App Store in-app prices (CONFIRMED 2026-10-19, not re-read): US Plus $19.00, Pro $39.00, Max $99.00; UK Plus £19.00, Pro £39.00, Max £99.00. API kimi-k3 unchanged (not re-read today): $3.00 in / $0.30 cached / $15.00 out per M; cache write $3.00 (5 min) or $6.00 (1 h). as of 2026-10-20
Scout, alternatives.json v3.72026-10-20date written in Scout's pricing text
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/tweet-triage-2026-10-14-scout.md (section for Kimi membership (Kimi Code, K3)).
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Merged from research/tweet-triage-2026-10-14-scout.md by merge_triage_table.py. [2026-10-18 Scout: Kimi Code community guidelines: "Kimi Code subscriptions are for personal interactive use only"; scripted batch execution "goes beyond normal use"; account sharing via reverse proxy is a violation; enforcement = concurrency limit, then possible suspension. Claude Code setup: ANTHROPIC_BASE_URL=https://api.kimi.ai/coding/, model k3 / k3-256k / k3[1m] (official docs). Pilot = one person interactive seat (Adam approved 10-18), not claude -p batch; batch work belongs on the pay-per-token API.] [2026-10-19T22:08Z cloud refresh: Merged from research/tweet-triage-2026-10-14-scout.md by merge_triage_table.py. [2026-10-18 Scout: Kimi Code community guidelines: "Kimi Code subscriptions are for personal interactive use only"; scripted batch execution "goes beyond normal use"; account sharing via reverse proxy is a violation; enf [2026-10-19 Scout T2 price/licence check: Plan names changed. The old tiers (stored as $19/$39/$99/$199) are now called legacy plans; existing subscribers keep them. New plans with Kimi Code (Plus and up) have a rolling 5-hour limit and no weekly limit. I could not find official USD prices for Go/Plus/Pro/Max/Ultra.]] [2026-10-19T23:00Z cloud refresh: 2026-10-19 session 2: USD prices for the new plan names found on the official App Store listing (apps.apple.com/us/app/id6474233312, seller Beijing Moonshot Technology Co., Ltd). The same list still sells legacy names (Moderato $19, Allegretto $39, Allegro $99, Vivace $199). App Store prices may differ from web checkout. kimi.com/en/membership/pricing and kimi.com/llms.txt return the SPA shell with no prices.] [2026-10-20T01:33Z cloud refresh: type None -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Kimi user agreement (prohibits reverse engineering; closed service). self_hostable, api_or_webhooks, data_export left out: the membership pricing page does not document them.] [2026-10-23T13:15Z cloud refresh: 2026-10-20: Ultra USD now confirmed ($199 monthly / $1,908 a year). Go has no USD price: it is not on the international page.] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): api_or_webhooks true: Kimi Code docs show membership API keys and an Anthropic-compatible endpoint (https://api.kimi.ai/coding/). self_hostable false from the Kimi API help FAQ (cloud API only, no local private deployment; private deployment via sales). The open K3 weights are a separate thing; the membership service itself is hosted. data_export left out (no export doc).]
Ultramarine 1 Flash (Paper Instruments)
commercialnot runnable on gaents yet
See research/tweet-triage-2026-10-18-scout.md (section for Ultramarine 1 Flash (Paper Instruments)). (Scout's description)
Ultramarine 1 Flash License (custom; HF tag license:other). Free for anyone who is not a 'Model Provider' (paid serverless or dedicated hosting for third parties); Model Providers need a separate agreement with Paper. Adds a prohibited-uses list. Base withheld MIT notice kept in NOTICE. Weights gated on HF (auto-approve form).
Paper API $0.15 in / $0.03 cached / $0.50 out per M (unchanged, same as GLM-5.3 Flash); Fireworks serverless (US) $0.225 / $0.045 / $0.75 per M; Baseten dedicated (no price on page); weights free to self-host for non-Model-Providers as of 2026-10-30
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/tweet-triage-2026-10-18-scout.md (section for Ultramarine 1 Flash (Paper Instruments)).
Credited to Scout (2026-10-30T21:49Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Merged from research/tweet-triage-2026-10-18-scout.md by merge_triage_table.py. [2026-10-20T01:33Z cloud refresh: type None -> commercial (no licence published; product today is a paid API; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from product page; licence_class and self_hostable left out (page says 'Model weights coming soon', no licence published); data_export n/a (inference API).] [2026-10-30T21:49Z cloud refresh: Re-check 2026-10-30 (bead gb4): open weights are out at huggingface.co/paperinstruments/Ultramarine-1-Flash (repo created 2026-10-22, last modified 2026-10-28; gated 'auto'; 105 files, ~331 GB; safetensors 321.3B params, mostly FP8 E4M3 + BF16; README: 320B total / 18B active, 1M context; LoRA rank-64 fine-tune of zai-org/withheld; served with SGLang v0.5.22 on 4x B300). Licence: custom 'Ultramarine 1 Flash License' (not MIT). licence_class source-available is agent-chose (Model-Provider exclusion + use restrictions, same rule as the DCG rider row); type left 'commercial' for Scout to decide. Also on Fireworks serverless (created 27 Oct 2026) and Baseten (Baseten page says 'Modified MIT License'; the HF LICENSE.md does not use that name). Harness/training code github.com/paper-instruments/feather-train: MIT, 226 stars (2026-10-30). Independent benchmarks: none found (no Artificial Analysis page, HTTP 404; not listed on withheld). Details: research/ultramarine-recheck-2026-10-30.md]
HarnessRouter (Community Edition)
open-corenot runnable on gaents yet
Self-hosted gateway implementing the Unified Harness Protocol (UHP): one OpenAI Responses-compatible API that runs tasks on 31 agent harnesses (Codex, Claude Code, Hermes, Pi, DeepSeek Harness, OpenCode, Gemini CLI, Aider, OpenHands, Grok Build ...) in their native runtimes, chosen per call with metadata.harness_id. Sessions, streaming, files/artifacts, cancellation, structured failures. (Scout's description)
Self-host free (Docker). HarnessRouter Cloud: metered by agent-work seconds (wait time $0.00); free plan keeps task records 7 days; tracing $50/mo add-on except on Scale. Model costs are your own provider keys. as of: no date recorded for this field
Scout, alternatives.json v3.72026-11-17T22:13ZScout patch_history.at; no date recorded for this field itself
GitHub stars
4360 as of: no date recorded for this field
Scout, alternatives.json v3.72026-11-17T22:13ZScout patch_history.at; no date recorded for this field itself
Scout local research pass (2026-11-09-tweet-triage-1103.json); re-verify price and licence before acting.
Limits for us
Needs a provider API key per harness ('No bundled model or trial key'); our Claude work runs on claude -p subscription logins and we hold no usable Anthropic API key, so the Claude Code harness has no credential to use here today. Container runs as root to manage per-session users. YC company (Epsilla) with a paid Cloud.
Lock-in risk
low
Integration hours (estimate)
8
Feature overlap (Scout's estimate)
0.4
Credited to Scout (2026-11-17T22:13Z, Scout patch_history.at). gaents does not score items.
https://github.com/HarnessRouter/harnessrouter (read 2026-11-09): “31 harness bases today: Codex, Claude Code, Hermes, Pi, DeepSeek Harness, OpenCode, Kilo Code, Qwen Code, Gemini CLI ... You need: Docker · About 4 GB of disk · A provider API key ... No bundled model or trial key.”
https://harnessrouter.ai/pricing (read 2026-11-09): “Waiting for a model response ... are Free wait time, shown on every receipt at $0.00 ... on the free plan that record is kept for 7 days ... available on other plans as a $50 per month add-on.”
Scout's notes
Scout tweet triage 2026-11-03 (research/tweet-triage-2026-11-03-scout.md) item 1. Brian Roemmele's 'one coordinator + HarnessRouter specialists' pattern. Tweet says 'roughly two dozen other harnesses'; README lists 31 harness bases (claim holds). Overlap with our own dispatcher/comms + claude -p workers + NTM: we route work between agents by role, not by harness per call.
AMD Token Lab (AMD AI Developer Program)
commercialnot runnable on gaents yet
AMD's free developer program; member benefits include 'instant API access through Token Lab' (https://notebooks.amd.com/tokenlab), an API for hosted open models. UNVERIFIED BEHIND LOGIN: we did not sign up, so the model list, quotas, card requirement and endpoint type are unknown. (Scout's description)
Free to join the AMD AI Developer Program. Token Lab quotas UNVERIFIED (behind login); a secondary source citing AMD AI DevDay 2026 says 25M free API tokens at launch. A separate $100 AMD Developer Cloud credit goes through DigitalOcean billing (third-party write-up). as of: no date recorded for this fieldunverified in Scout's data
Scout, alternatives.json v3.72026-11-17Scout evidence.read_at; no date recorded for this field itself
Scout tweet triage 2026-11-17 (Adam's batch; reply COO/requests/tweet-triage-2026-11-17/reply-Scout.md); re-verify price and licence before acting.
Limits for us
Unverified behind login. The tweet's model list (DeepSeek-V4-Flash, withheld, Qwen3.8, Gemma-4-31B), 'no card' and 'works with Claude Code's Anthropic endpoint' are unverified. Probably a capped token grant (25M tokens ~ 10-20k of our lean fact calls), not a lasting supply.
Lock-in risk
low
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.3
Credited to Scout (2026-11-17, Scout evidence.read_at). gaents does not score items.
Tweet triage 2026-11-17 #4. Details: /home/a/Projects/COO/requests/tweet-triage-2026-11-17/reply-Scout.md #4. Verdict-type note: unverified behind login (we did not sign up). Similar free tiers already in this category: NVIDIA build.nvidia.com free tier, withheld, freellmapi. Possible use: non-Claude fallback for lean extraction (facts runner) via a 50-row A/B against Haiku as in data/model-ab-2026-11-09/. Needs one account by the org keys owner (Forge or Hustle), if Adam agrees.