groundy
Models & Research

Claude Code or Cursor on Inkling 2 Small: What to Verify Before You Route

Inkling 2 Small coding metrics are vendor-reported, so teams must verify served context, costs, and agent behavior in Claude Code, Codex, or Cursor before routing real work.

Published 9 references
A skeptical green resin dinosaur reaches one yellow-tipped forefoot onto an irregular low step, with an open resin door behind it. Bubbles, scuffs and hard shadows texture the warm ivory scene.
On this page11 sections

Route a real repository to Inkling 2 Small only after it passes an evaluation on your own tasks. Vercel’s November 12, 2026 changelog makes the model easy to select in Claude Code, Codex, and Cursor, but every coding performance figure published for the model is vendor-reported. As of November 15, 2026, the evaluations in circulation come from Thinking Machines itself or from distributors repeating its numbers. That absence is not proof that no independent evaluation exists anywhere; it does mean the launch numbers cannot be your routing gate.

The distribution terms lower a different barrier. Vercel says AI Gateway passes through provider pricing without markup and charges no platform fee on inference, including Bring Your Own Key requests. That makes the connection cheap to try, not free to run. You still pay for tokens, and the important costs may come from testing a new route, checking its limits, and diagnosing agent behavior that differs from the launch configuration.

The practical shift is therefore from “Can we connect this model?” to “Can we establish that it works in our harness?” The first question now has a documented answer. The second is your team’s responsibility.

What the Vercel listing establishes, and what it leaves open

The documented setup is unusually short. Install the latest Vercel CLI, run:

Terminal window
npx vercel ai-gateway setup

Then select thinkingmachines/inkling-2-small in the agent. Vercel’s changelog explicitly names Claude Code, Codex, and Cursor.

Vercel describes Inkling 2 Small as accepting text and image inputs and producing text output, with support for tool use during extended coding tasks. It also says the model “can use screenshots to check the output of the code it writes and guide the next revision.” The first-party model card confirms the text-and-image input and text-output specification.

Those are compatibility and capability claims from the vendors, not evidence that every agent preserves every behavior. In particular, the listing does not establish that Claude Code, Codex, and Cursor all pass screenshots back to the model in the same way, expose equivalent tools, or use the same reasoning settings.

As of November 15, 2026, the changelog does not list three pieces of decision information:

  • The model’s exact per-token charge on the gateway route.
  • The context window actually served through that route.
  • Model-specific rate limits.

Those are open questions, not inferred negatives. Vercel may document them elsewhere, providers may expose them in account dashboards, and terms may change. Verify the route you will actually use rather than filling the gaps with another provider’s specifications.

Advertised context is not necessarily served context

The Inkling 2 Small model card advertises a context window of up to 1M tokens. It lists a November 5, 2026 release, Apache 2.0 licensing, 594B total parameters with 17B active per token, and MXFP8 and NVFP4 checkpoints. Together AI’s model page also lists a 1M-token context window.

The counterevidence comes with a provenance warning. The unofficial tracker inkling2small.com reports advertised context of 1M tokens but served context of 64K or 256K tokens for variants on Tinker. That is a lead to verify, not an established measurement of Vercel’s route or of every Tinker deployment. The tracker’s reliability is not established.

The distinction matters because “the model supports 1M” and “this endpoint serves 1M” are different claims. A provider can expose a smaller context than the model card’s maximum. If your expected repository task requires more input than the served route accepts, the advertised maximum does not help that request.

For a large-repository workflow, verify both the configured limit and the behavior near it. Use a representative long-input task and record whether the request succeeds, what the endpoint reports, and whether the agent still completes the task correctly. Do not assume that a successful short coding task validates the 1M claim. Nor should the unofficial 64K/256K figures be treated as proof that the gateway has the same limits.

Check the model name before comparing specifications. First-generation Inkling-Small, released July 30, 2026, is a different model: 276B total parameters, 12B active, with text, image, and audio input. Inkling 2 Small is 594B/17B and accepts text and images. Their prices, context claims, and benchmark results should not be mixed.

Cost depends on the route, not just the model

Vercel’s no-markup statement answers whether the gateway adds a platform surcharge. It does not identify the underlying provider price for your Inkling 2 Small requests. The available third-party prices are useful reference points, but neither should be treated as a confirmed gateway invoice.

RouteDocumented cost or requirementContext evidenceWhat remains to verify
Vercel AI GatewayProvider pricing with no markup and no platform fee on inference, including BYOKNot specified in the changelog as of November 15, 2026Exact per-token invoice, served context, and model-specific rate limits
Together AI$0.30 per 1M input tokens, $0.03 cached, and $1.00 per 1M output tokensTogether lists 1M tokensWhether these direct-provider terms match any route you actually select
TinkerThe unofficial tracker lists $0.30 prefill, $0.03 cached prefill, and $1.00 sample per 1M tokensThe tracker reports 64K/256K served variantsFirst-party confirmation of current prices, variants, and limits
Self-hostingThe model card sets a minimum of 500GB aggregate VRAM for the main NVFP4 checkpoint, such as 2 NVIDIA B300 or 4 NVIDIA H200 GPUs; MXFP8 requires at least 800GBControlled by your deployment, subject to the model’s supported maximumHardware, serving, operations, and evaluation costs

The hardware floor prevents a common misreading of “17B active.” The active-parameter count does not remove the model card’s listed memory requirement for the full checkpoint. Self-hosting is an exit option, not the cheapest way to run a quick trial.

Plan allowances are separate from inference fees. Vercel’s pricing page lists BYOK among AI Gateway features and shows 500 included Token Requests per month on Hobby. That is an included plan allowance, not evidence of an Inkling-specific rate limit. Before an agent workload, check your plan’s current allowance, overage terms, and any provider-side limits.

Read the coding scores with their settings attached

Together AI’s benchmark compilation includes a vendor-reported 82.2% on SWE-bench Verified and 25.4% on FrontierCode v2.0, among other results. Together explicitly discloses that the results were “reported by Thinking Machines Lab at effort 0.99 (xhigh) and temperature 1.0,” using externally reported evaluations where available.

That disclosure is more useful for routing than the headline percentages alone. A result at effort 0.99 is not a promise about your agent’s default. Vercel’s setup instructions do not specify that Claude Code, Codex, or Cursor will use the same effort or temperature. Check what your harness sends, and run your comparison at settings you intend to pay for and operate.

The harness also forms part of the result. Together reports a vendor-reported 85.1% on Terminal-Bench 2.1 specifically with the Pi harness. The unofficial tracker, labeling its table as reported on the official model card, lists 81.3% for Terminal Bench v2.1 with terminus v2. These are different harnesses and different scores within the same benchmark family. They should not be collapsed into one context-free terminal-coding number.

For Claude Code specifically, the tracker lists a vendor-sourced 38.8% on SWE-Marathon v1.1 run with Inkling 2 Small in Claude Code, and the first-party model card reports the same figure. It is the only coding-benchmark row on the model card that names the Claude Code harness; the other coding rows use harnesses such as stirrup, mini-swe-agent, terminus, and openhands. The model card does carry one more Claude Code row outside coding: PostTrainBench v1.2 at 31.9%, filed under Autoresearch. None of these results is independent replication, and none establishes comparable performance in Codex or Cursor.

The reasonable conclusion is narrow: the model has published vendor results worth investigating, but the numbers do not predict your repository pass rate. Model, effort, temperature, tools, harness, and task suite all belong in the comparison.

What changes between Claude Code, Codex, and Cursor?

On the gateway, the documented connection cost is the same for all three: one setup command, provider pricing passed through, and no platform fee on inference. The exact token invoice remains unresolved. The important differences are therefore in the harness and the evidence available for each pairing.

AgentWhat the sources establishPublished model-specific evidenceVerify before routing repository work
Claude CodeVercel documents the one-command routeVendor-sourced SWE-Marathon v1.1 result of 38.8% in Claude Code, surfaced by the trackerYour effort default, tool behavior, long-context handling, screenshot loop if used, and pass rate against your current baseline
CodexThe same Vercel setup and model selectionNo published Codex-specific Inkling 2 Small evaluationThe same operational checks in Codex itself; do not substitute the Claude Code result
CursorThe same Vercel setup and model selectionNo published Cursor-specific Inkling 2 Small evaluationWhether Cursor’s actual workflow exercises the vendor-claimed tool and visual behaviors, plus context, cost, and task correctness

Nothing in this matrix says one agent is the better choice. The evidence cannot settle that comparison. It supports testing each candidate pairing rather than treating “works in Claude Code, Codex, and Cursor” as proof of equivalent behavior.

There is also a configuration detail worth checking manually. Vercel gives the model string as thinkingmachines/inkling-2-small, while Together uses thinkingmachines/Inkling-2-Small. Preserve the exact string required by the selected provider instead of assuming identifiers are case-insensitive.

Expect integration and configuration work, not just model errors

An independent study, “Engineering Pitfalls in AI Coding Tools”, examined 3,864 user-reported issues across Claude Code, Gemini CLI, and Codex CLI. The authors report that API and integration errors accounted for 21.4% of categorized root causes, while configuration and setup accounted for 15.9%.

Those percentages are shares of root causes in that issue dataset. They are not a predicted failure probability for Inkling 2 Small, and the study does not test this model or Cursor. Its relevance is the failure profile: changing a model route introduces API and configuration surfaces whose behavior can be mistaken for weak model quality.

Keep those failures visible in your evaluation logs. Record setup failures, tool-call errors, rejected or over-context requests, and incomplete screenshot loops separately from wrong code. Otherwise, a routing problem can depress a task score without telling you whether the model or the integration caused it.

A workable evaluation costs less than an unexamined migration

A second independent preprint, “Should you use Claude to review Codex or vice versa?”, provides a cost template rather than evidence about Inkling. Using Claude Opus 4.7 and Codex GPT-5.5, the authors report that reproducing six writer-reviewer conditions across 116 complete tasks cost roughly $225 in API calls, or $1.94 per task across six conditions.

That is not a quote for your Inkling evaluation. Token use, task length, provider prices, and the number of conditions will differ. It does show that a structured comparison can have a bounded, measurable cost rather than requiring an open-ended pilot.

The same paper supplies a warning about cross-model review. The authors report that Claude review raised Codex draft performance from 71.6% to 89.7% in their setup, while the reverse pairing hurt. Pairing effects were asymmetric. Do not assume that adding Inkling as a writer or reviewer will help merely because another model pairing did.

A useful pre-route evaluation can stay compact:

  1. Freeze the task set. Choose representative tasks with tests or another predetermined correctness check.
  2. Keep the harness fixed. Compare your current route and Inkling 2 Small in the same agent, with the same tools and task instructions.
  3. Record the settings. Capture the exact model string, reasoning effort, temperature where exposed, served context, and provider route.
  4. Measure operational failures separately. Track integration, configuration, context, and tool errors alongside code correctness.
  5. Measure actual consumption. Use returned usage and billing records to calculate route-specific cost rather than substituting Together or Tinker prices.
  6. Test the features you intend to use. If screenshot-driven revision matters, verify that the complete agent loop works. A general vision-input specification is not enough.

Exit options are real, but they do not settle performance

Thinking Machines says the weights became downloadable on HuggingFace on November 12, 2026. Its launch update also lists API availability through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare. Together with the Apache 2.0 license and documented self-hosting checkpoints, that gives a team alternatives to a single gateway endpoint.

Those alternatives reduce dependency on one distribution route. They do not establish identical context windows, prices, reasoning defaults, latency, or agent behavior across providers. An exit ramp still needs its own verification.

My routing call is to treat Inkling 2 Small as a candidate for a controlled trial, not a default for real repositories. Proceed when the exact route’s context, settings, request limits, and costs are known enough for your workload, and when your fixed task evaluation beats or acceptably matches the current baseline. For a large repository, unresolved served context is a blocking question. For tool-heavy or visual development, an unverified agent loop is the blocker.

The open research question remains independent evaluation. As of November 15, 2026, everything published on the model’s coding performance is vendor-reported, so the evidence supports a verification process and route-specific cost comparison, not a capability verdict.

Frequently Asked Questions

What is the minimum hardware requirement for self-hosting Inkling 2 Small?

The model card sets a minimum of 500GB aggregate VRAM for the main NVFP4 checkpoint, such as 2 NVIDIA B300 or 4 NVIDIA H200 GPUs; MXFP8 requires at least 800GB

What is the exact model string to use for Inkling 2 Small on Vercel?

Vercel gives the model string as thinkingmachines/inkling-2-small

References

Follow the links in the article for context. The supporting material is collected here for further reading.

  1. Vercel's November 12, 2026 changelogvercel.comAccessed
  2. first-party model cardthinkingmachines.aiAccessed
  3. Together AI's model pagetogether.aiAccessed
  4. inkling2small.cominkling2small.comAccessed
  5. Inkling-Smallthinkingmachines.aiAccessed
  6. Vercel's pricing pagevercel.comAccessed
  7. Engineering Pitfalls in AI Coding Toolsarxiv.orgAccessed
  8. Should you use Claude to review Codex or vice versa?arxiv.orgAccessed
  9. launch updatethinkingmachines.aiAccessed

Join the discussion

Share a useful perspective or ask a question about this article.

Discussion guidelinesComments privacy