Inkling 2 Small vs Inkling Small: Pricing, Capability and Which Fits Your Work

Short answer

Thinking Machines shipped exactly one model this round: Inkling 2 Small (inkling-2-small), released November 5, 2026, the first of the Inkling 2 generation. The only same-vendor predecessor the release material lets us confirm is Inkling Small.

So "which of the vendor's models should I pick" has a simpler answer than you might expect:

• New project, new pipeline → inkling-2-small. It scores 31 on the AA Intelligence Index where the previous generation scored 13, and output costs $1.00 against the older model's $1.20. It wins on both capability and price, and the release material offers no reason to stay on the old one.
• Existing pipeline pinned to the previous generation → price out the migration first. Short of that kind of legacy constraint, there is nothing to agonize over between the two.
• The decision actually worth your time is whether to use it at all, not which generation. It ranks near the top on agentic coding and vision, does badly on terminal tasks, and is slow and expensive per task.

inkling-2-small is not on TeamoRouter yet. Once it lands, change the model field to inkling-2-small on the key you already have; the base URL stays put. Until then: TeamoRouter carries no Thinking Machines models at all, so there is no same-vendor substitute. We are not going to hand you another lab's model as a stand-in — the capabilities, behaviour and pricing would not line up. If you need it today, Tinker is the only place to get it.

Two generations, one table

Every "not disclosed" is literal. The release material says exactly two things about the previous generation — an AA Intelligence Index of 13 and an output price of $1.20 — and we are not filling in the rest on its behalf.

What actually changed between them

Capability: not a tune-up, more than a doubling. 13 to 31 on the same Artificial Analysis Intelligence Index. A jump of that size is not what re-tuning a model looks like; the training path changed.

Price: generations usually get more expensive. This one went the other way. Output fell from $1.20 to $1.00. Only the new model's input price was published at $0.30, so there is nothing to compare on that side.

Training: two runs merged became one run. The previous Inkling ran separate reasoning and agentic RL passes and merged them by distillation. This one is a single joint RL run — 1,075 steps, more than 17M rollouts. What that buys you in practice is that "reasoning well" and "acting well" are no longer trained in separate rooms. Agent work needs both step-by-step planning and tool calls in the same turn, and that is where joint training pays off.

Positioning: much clearer this time. Thinking Machines describes Inkling 2 Small as a multimodal model for agentic coding, vision and knowledge work: text and images in, text out, audio explicitly unsupported. The release material does not restate how the previous generation was positioned.

Where it shines, where it falls down

Every score below is from Thinking Machines' published benchmark table, run at effort 0.99 (xhigh) and temperature 1.0 — the "think as hard as you can" setting. Day-to-day calls will not need that; it is simply the setting where the gaps open up widest.

Strengths:

• Code changes: 86.3% on SWE-Bench Pro v2 (Hard), second in the lab's own table only to Kimi K3.
• Reading screens and images: PerceptionBench 64.7%, against 58.5% for Kimi K3; DesignArena Website 1,340, third on the November 5 leaderboard snapshot.
• Enterprise workflow tasks: EnterpriseBench-CoreCraft 71.8% against 46.7% for GLM-5.3-Flash in the same table; JobBench 70.5%.
• Low-level optimisation: 54.25 of 69 tasks solved on average on the PMPP-Hard GPU kernel benchmark.

Weaknesses:

• Terminal and command-line work: 13.1% on Terminal-Bench 4.0. The same material puts GLM-5.3-Flash at 32.8% and DeepSeek-V4.1-Flash-1017 at 33.8%.
• General engineering tasks: 72.1% on DeepSWE v1.1, against 77.0% for DeepSeek-V4.1-Flash-1017.
• Knowledge-work scoring: 1,630 on GDPVal-AA v2.1, against 1,746 for DeepSeek.
• Speed: AA measures 21.5 minutes per index task — well behind DeepSeek's 13.3 minutes, ahead of GLM's 37.7. The lab's own summary is "slow and verbose".
• Audio: unsupported. Text and images in, nothing else.

In one line: it reads like a colleague who is very good at writing code and very good at reading a screen, and who should not be left alone with a shell.

A cheap unit price is not a cheap task

This is the part people skip.

Per token, its output is cheaper than DeepSeek's. Per task, it costs nearly twice as much. The reason is straightforward: it spends more thinking tokens per turn and takes longer to finish. Which means:

• For quotes and budgets, use cost per task, not price per token. The gap between $1.18 and $0.60 is real once you are running thousands of tasks.
• For high-volume, short calls, per-token price is the number that matters.
• GLM is cheapest on both counts, but 37.7 minutes a task is a different kind of price — fine for offline batch work, wrong for anything interactive.

A practical order of decisions

1. Got a pipeline tied to the previous Inkling? Move it to inkling-2-small. Double the capability and cheaper output — there is no case for staying.
2. New work, no strong opinion yet? Start on inkling-2-small at the default reasoning effort and run a batch of your real requests. Do not open with xhigh; that is an evaluation setting, and it is slow and expensive.
3. Mostly terminals, scripts and unattended command-line work? Do not count on it. 13.1% is published, and the retries in that kind of work will eat the unit-price advantage.
4. Quoting per task? Start at $1.18 and add your own retry rate on top.
5. Audio in the pipeline? Use something else. It does not do audio.

Where it sits in its price class

The lab's claim is fourth of 29 models in its price band on the AA index. Item by item:

Three things to take from it:

• It does not top the index. 31 is below DeepSeek's 33 and above GLM's 28. Its wins are in specific rows — vision and agentic coding — not in the aggregate.
• Parameter count is not efficiency. 594B total with only 17B active is the same idea both competitors use. What separates them is how many thinking tokens each burns per turn, and that is what drives cost per task.
• The licence differs. Both competitors are MIT today; Inkling 2 Small is closed at launch, with Apache 2.0 planned after external safety testing and red-teaming. Real self-hosting waits on the weights.

Calling it on TeamoRouter

Not yet available, so this is what it will look like once it lands:

from openai import OpenAI

client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
    model="inkling-2-small",
    messages=[{"role": "user", "content": "Rewrite the state management in this React component"}],
)
print(resp.choices[0].message.content)

inkling-2-small is not on TeamoRouter yet. Once it is live, the same key and model: inkling-2-small is all you change; there is no same-vendor substitute in the meantime, so if you need it now, Tinker is the only option. Pricing once it is listed will be whatever the console says.

How to pay

No subscription: top up your balance in the console from $5 and each request is billed by the tokens it uses. Cards are supported, and so are USDT, USDC and ETH. The same key also calls Claude and GPT models.

New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.

FAQ

Are there other models from the same vendor to choose from?

The only same-vendor model the release material supports is the previous Inkling Small. Inkling 2 Small is the first of generation two; there are no sibling models.

How big is the gap between the generations?

31 against 13 on the AA Intelligence Index, and $1.00 against $1.20 on output. The previous model's other specs were not published.

Why is the newer generation cheaper on output?

The material gives the numbers and no explanation. All we can confirm is that $1.00 is lower than the earlier $1.20.

What is it best at?

Agentic coding (86.3% on SWE-Bench Pro v2 Hard), reading images and interfaces (64.7% on PerceptionBench), enterprise workflow tasks (71.8% on EnterpriseBench-CoreCraft, 70.5% on JobBench), and low-level work like GPU kernels.

What is it worst at?

Terminal and command-line tasks — 13.1% on Terminal-Bench 4.0. It also does not handle audio.

Is there a free tier?

The release material mentions no free allowance, only Tinker's per-token pricing.

Can I self-host it?

Not yet. Weights are closed at launch, with Apache 2.0 planned after external safety testing and red-teaming. Hardware requirements for self-hosting were not disclosed.

Can I use it on TeamoRouter now?

Not yet — inkling-2-small has not entered our catalog. Once it does, the same key and a new model ID is all it takes.

Related reading

• Inkling 2 Small API is out: model ID, Tinker pricing and both ways to call it
• Using Inkling 2 Small without a Tinker account: what you can do without an account or a card
• The release write-up: source for the specs, scores and prices