A headline reading "Thinky Machines Makes Inkling MoE Models Free on OpenRouter With 1M Context Window" is trending on AI news aggregators today, August 22, 2026. Two things are worth clearing up before treating it as fresh news: "Thinky Machines" is a garbled rendering of Thinking Machines Lab, and the "free on OpenRouter" part is not a new price change — it is the same standing offer that has existed since Inkling's July 17, 2026 launch.
That doesn't make the story worthless. It makes it a fine-print story — exactly the kind explainx.ai has been tracking all week with OpenRouter's GPT-5.6 Sol "50% off" promo and the Ox Alpha stealth model's "free for now" caveat. A free-sounding router listing is real, but the terms attached to "free" are usually the part that determines whether you can actually build on it.
TL;DR
| Question | Direct answer |
|---|---|
| Who is "Thinky Machines"? | A garbled aggregator rendering of Thinking Machines Lab, Mira Murati's AI company. Likely traces to the lab's own X handle, @thinkymachines. No separate company exists under that name. |
| What is Inkling? | Thinking Machines Lab's first open-weights model — a 975B-parameter mixture-of-experts (MoE), 41B active parameters per token, released July 15–17, 2026. Full coverage: explainx.ai's Inkling guide. |
| Is Inkling free on OpenRouter? | A free endpoint exists, but it's restricted to agentic harnesses, logs prompts and outputs for Thinking Machines Lab's own model improvement, and disallows confidential/personal data. It is not an unconditional free tier. |
| Is this a new announcement? | No. The free endpoint has existed since Inkling's July 17, 2026 launch on OpenRouter. This is a five-week-old offering surfacing again on an aggregator, not a fresh price change. |
| What does the 1M context window actually get you? | 1,048,576 input tokens, up to 262,144 completion tokens — enough for a full mid-size codebase, a long document set, or hours of agent transcript in a single context, per Inkling's original spec. |
| What does paid Inkling cost? | $0.95 / 1M input tokens, $4.05 / 1M output tokens, $0.16 / 1M cached-read tokens on OpenRouter. |
| Should I build on the free tier? | For evaluation and non-sensitive agent testing, yes. For production or anything with customer data, budget for the paid rate or a hosted-provider key instead. |
Identity check: "Thinky Machines" is Thinking Machines Lab
Before anything else, the naming confusion needs to be resolved directly, because it's the kind of garbling that spreads fastest on aggregator feeds that strip context from a headline.
Thinking Machines Lab is the AI company co-founded by former OpenAI CTO Mira Murati, alongside former OpenAI researcher John Schulman, built around a stated mission of AI that extends human judgment rather than replacing it — laid out in the lab's "Future Worth Building Is Human" manifesto on July 10, 2026. explainx.ai has covered the lab's release cadence in depth: the manifesto itself, the Inkling launch, and the Inkling-Small preview.
The most likely source of "Thinky Machines" is OpenRouter's own launch-day post on X, which reads: "Inkling from @thinkymachines is live on OpenRouter. An open-weights MoE model with 975B total / 41B active parameters, 1M context, and controllable reasoning across text, images, and audio." Thinking Machines Lab's X handle is literally @thinkymachines — drop the "@" and a headline-generation pass, or an aggregator paraphrasing a tweet, plausibly renders that as "Thinky Machines." There is no evidence of any distinct company by that name; every specification in the trending headline (Inkling, MoE, OpenRouter, 1M context) matches Thinking Machines Lab's actual model exactly.
What Inkling actually is
Inkling is not a new model as of this week — it shipped July 15, 2026, with the OpenRouter endpoint live by July 17. The core specs, confirmed on Thinking Machines Lab's own model card and unchanged since launch:
| Spec | Value |
|---|---|
| Architecture | Mixture-of-experts, DeepSeek-V3-style routing — 256 routed experts + 2 shared experts, 6 active per token |
| Total parameters | 975B |
| Active parameters | 41B per token |
| Context window | 1,048,576 tokens (≈1M), up to 262,144 completion tokens |
| Modalities | Text, image, and audio input; text output |
| Pretraining | 45 trillion tokens across text, image, audio, and video |
| Positioning | Best open-weights base for customization (via Tinker fine-tuning) — explicitly not marketed as the top score on every leaderboard |
For the full benchmark tables — SWE-bench Verified, Terminal Bench 2.1, HLE, and how Inkling stacks against Kimi K2.6, GLM 5.2, Claude Fable 5, and GPT-5.6 Sol — see explainx.ai's dedicated Inkling coverage, which remains the definitive reference; this post focuses specifically on the OpenRouter free-endpoint mechanics that are driving today's trending headline.
The free endpoint: what "free" actually means
This is the section that matters most if the trending headline is why you're here. OpenRouter's model page for Inkling states the free-tier terms plainly, and they are meaningfully narrower than "free for everyone, for everything":
- Agentic-harness-only. OpenRouter's own language: "The free Inkling endpoint is only available for use with agentic harnesses." A general chat client hitting the free endpoint outside that use case is outside the terms as documented.
- Logged for model improvement. Prompts and outputs sent to the free endpoint are logged and used to improve Thinking Machines Lab's own models, products, and services — the tradeoff for zero cost is your traffic becoming training or evaluation signal.
- No confidential or personal data. The terms explicitly instruct against sending confidential or personal data through the free endpoint.
- Requires accepting the TML Free Research API Terms of Service — a separate agreement from OpenRouter's standard terms.
- Routed across multiple providers. Inkling on OpenRouter runs through DeepInfra, Together, and Baseten with automatic failover — standard OpenRouter routing behavior, not a dedicated Thinking Machines Lab endpoint.
None of that makes the free tier fake. It's a genuinely no-cost way to run a 975B-parameter MoE model without a GPU cluster or an API contract. But it is scoped, logged, and use-case-restricted — closer to a research/evaluation tier than a production giveaway, and reading it that way avoids the mistake of budgeting a real product around it.
How this compares to this week's other OpenRouter "free/discounted" stories
explainx.ai has flagged this same pattern twice already this week, and it's worth naming as a general rule rather than a one-off:
| Story | What the headline implies | What the terms actually say |
|---|---|---|
| GPT-5.6 Sol "50% off" | OpenAI cut Sol's price | A time-limited, non-BYOK OpenRouter promo — OpenAI's own price card is unchanged |
| Ox Alpha "free" stealth model | A frontier-class model, free indefinitely | Free "for now," with an explicit open question about what happens once Stripe's acquisition of OpenRouter closes |
| Inkling "free" (this post) | A 975B MoE model, unconditionally free | Free, but scoped to agentic harnesses only, logged for training, and has been standing since July — not new |
The throughline: a router listing badge is a real fact, but it is not the whole contract. Read the linked terms page before you route production traffic through any "free" label on OpenRouter.
What the 1M context window is actually useful for
A 1M-token window is large enough to matter for specific workloads, not a universal upgrade:
- Whole small-to-mid codebases in a single context, without chunking or retrieval — useful for agentic code review or refactor planning inside a harness like OpenCode.
- Long agent transcripts — hours of tool-call history in one pass, relevant to the "agentic harnesses only" restriction on the free tier, since long-running agent loops are exactly what benefit from not truncating context.
- Document-set analysis — contracts, research corpora, or logs too large for a 128K–200K window without splitting.
It is not, on its own, a reason to prefer Inkling over a smaller-context model for short, single-turn tasks — the effective cost and latency of filling a 1M window on the paid tier ($0.95/$4.05 per 1M tokens) adds up fast if you don't need the length. Inkling's own controllable thinking-effort dial is the more relevant cost lever for most day-to-day use.
How to actually try it
# OpenRouter API — free endpoint (agentic harness use case)
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "thinkingmachines/inkling:free",
"messages": [{"role": "user", "content": "Summarize this 200-page contract."}]
}'
Practical guidance:
- Confirm your harness qualifies — the free endpoint's terms name agentic harnesses specifically; a plain chatbot integration may fall outside intended use.
- Never route confidential or customer data through the free endpoint — it's logged by design.
- Budget the paid rate for anything production-facing. At $0.95/$4.05 per 1M tokens, a full 1M-token context fill on input alone runs close to $1 per call before output — model that against your workload before committing.
- Compare against Kimi K2.6 and GLM 5.2 on your own tickets before switching defaults — see explainx.ai's Kimi K2.7-Code guide for a coding-specialist alternative with different tradeoffs.
Bottom line
"Thinky Machines Makes Inkling MoE Models Free on OpenRouter" is, once decoded, an accurate but stale-sounding way to describe a real thing: Thinking Machines Lab's Inkling has carried a scoped, logged, agentic-harness-only free endpoint on OpenRouter since mid-July 2026, alongside a paid tier and a genuine 1M-token context window. Nothing changed today except the headline resurfacing on an aggregator. If you're evaluating Inkling this week, the free tier is a legitimate way to kick the tires — just read the terms, keep sensitive data off it, and budget the paid rate before anything reaches production.
Related on explainx.ai
- Inkling: Thinking Machines Lab open-weights MoE — full specs and benchmarks
- Inkling-Small — 12B-active preview
- Thinking Machines Lab manifesto — The Future Worth Building Is Human
- GPT-5.6 Sol is not 50% cheaper — OpenRouter promo, not a price cut
- Ox Alpha: the free stealth model on OpenRouter
- Stripe acquires OpenRouter for $7 billion — what builders should know
- What is OpenRouter? Enterprise guide
- Kimi K2.7-Code — open coding specialist
- How to run open-source models locally with OpenCode
- AI token pricing, explained
Official: Inkling model card — Thinking Machines Lab · Inkling on OpenRouter · Introducing Inkling
This post decodes a trending aggregator headline as of August 22, 2026. Free-tier terms, pricing, and routing details reflect OpenRouter's live model page at time of writing and are subject to change without notice — verify current terms on OpenRouter before routing production traffic. "Thinky Machines" identification is based on specification matching and the likely @thinkymachines handle source; no statement from an entity using that exact name was found.