A small Dutch site called notanaiapp.com ran the most useful AI coding test I have seen this month, and the result is uncomfortable for anyone who has picked a side in the editor wars. On October 10, the testers gave Cursor and Claude Code the same three coding jobs, ran each job three times, and checked the output against 186 hidden tests. Their move that made it interesting: they forced both tools to use the same model, Claude Sonnet 5.5.
The result: both tools passed all 186 checks, every time. Claude Code averaged about 36 seconds per job, Cursor about 33. Statistically, a tie.
Then comes the number that should reframe how you evaluate these tools. Claude Code sent roughly 269,000 tokens to the model per job. Cursor sent about 92,000. Same model, same tasks, same pass rate, and the tool wrapped around the model changed the consumption by a factor of 2.9.
Most of that difference is repeated context: the same file contents read back into the model over and over across agent steps. At Anthropic's API prices, that works out to about $0.14 of token consumption per job in Claude Code and $0.11 in Cursor. Not a huge gap in absolute terms, but the ratio matters when you scale it across a working day of agentic tasks.
There is a caveat, and the testers flagged it themselves: Cursor did not report cache-written tokens, so its real number is likely somewhat higher than 92,000. Even discounting for that, though, the shape of the finding holds. The tool shapes token consumption, not quality, when the model is held constant.
What the October 10 test actually measured
The methodology is worth respecting before drawing conclusions:
- Three coding jobs, run three times each, checked afterwards against 186 hidden tests.
- Same model, same effort setting. Both tools ran Claude Sonnet 5.5 at effort "medium", which is the fairest way to isolate tool effects from model effects.
- Result: Claude Code 186 of 186, Cursor 186 of 186, roughly 36 versus 33 seconds per job.
- Token consumption: about 269,000 tokens per job in Claude Code versus 92,000 in Cursor, a 2.9x gap.
They also ran two more configurations worth knowing about. Cursor's in-house model, Composer 2.5, scored 181 of 186, about 35 seconds per job. Cursor's free tier, which only lets you use the "Auto" model picker, also scored 181 of 186 but took about 76 seconds per job, roughly twice as slow as either tool on Sonnet 5.5.
The takeaway from those extra rows is that model choice still matters for speed and reliability. Sonnet 5.5 beat Composer 2.5 on the checks, and Auto was slower than either. It is specifically the tool wrapper that stopped mattering once the model was fixed.
Why this reframes the whole editor debate
Most editor comparisons, including some I have read and some published on big outlets, compare bundles: this editor with its default model versus that editor with its default model. That measures model choice as much as tool design. The October 10 test strips that out, and what is left is revealing.
Quality is model-bound. With the same model, both tools hit the same pass rate. The agent harness, the prompts, the loop structure, none of it changed whether the code was correct. If your main question is "will it produce working code", the honest answer is that the model decides, and the tool follows.
Efficiency is tool-bound. The 2.9x token gap is a tool-design difference. It reflects how much context each harness thinks it needs before acting, how aggressively it re-reads files, how it structures its agent loop. That is a real cost difference if you pay per token through an API, and a real speed difference on rate-limited plans.
Speed, at least here, was a wash. 36 versus 33 seconds is close enough that it should not drive your decision.
The conclusion the testers reached is that you should pick the tool that fits how you already work, not the one with the best benchmark chart. If you already pay for Claude, Claude Code is bundled in and did just as well here. If you want an editor with built-in AI and a model picker, Cursor is that. If you want free, Cursor's free tier works but is slower and slightly less reliable.
What changed since the test: the X Subscription
There is an update that landed after the test, on October 13, and it matters for anyone choosing this week. xAI now sells a bundle called The X Subscription that includes Cursor alongside Grok and X Premium:
- Plus, $8/month: Grok models only, inside Cursor.
- Premium, $30/month: bundles SuperGrok, Cursor Pro, and X Premium+ benefits. Outside models like Claude cost extra at API prices from this tier upward.
- Super, $100/month and Ultra, $200/month: higher limits across the bundle.
- Existing Cursor plans keep working for now.
This is a notable shift: a subscription best known for social media and chatbot access now ships with a serious coding editor attached. Whether that bundle is good value depends entirely on whether you wanted Grok models anyway. If your preference is Claude or GPT models in Cursor, the bundle's pricing gets less attractive at the Premium tier and above, since outside models cost extra on top of the subscription. There is also a geographic wrinkle: as of October 13, this tester in the Netherlands was still seeing Cursor's own standalone plans, so rollout may be uneven.
Where this fits in the October market picture
The same-model result lands in a month where the usage numbers finally got rigorous. Stack Overflow released its 2026 Developer Survey on October 6, covering 30,903 valid responses from 169 countries, and its coding-agent question is the cleanest adoption snapshot we have. Among respondents who use a listed AI coding agent, Claude Code led at 65.5 percent, GitHub Copilot followed at 58.7 percent, OpenAI Codex came in at 29.5 percent, and Cursor at 25.5 percent. VS Code, for what it is worth, still leads the editor category at 70.9 percent usage, with Cursor at 14.9 percent.
Those usage numbers and the same-model quality tie tell one coherent story: Claude Code leads on adoption and bundling, not on some intrinsic ability to write better code than Cursor can. Its advantage is distribution, a Claude subscription includes it, and the harness efficiency you pay for at API rates. Cursor's counter-offer is the editor experience, the model picker, and now the X Subscription bundle. Visual Studio Magazine's survey coverage puts the same numbers in context: VS Code leads editors while Claude Code leads coding agents, and the two rankings now move independently.
There is also an ownership angle worth one sentence when you pick this month: Cursor has been a wholly owned SpaceX subsidiary since the $60 billion acquisition closed on August 14, and its Ramp spending share has slid from 41 percent to 26 percent since, per LogRocket's October 2026 tool rankings, while Claude Code and OpenCode gained ground. Buying an editor is still, in a small way, choosing whose infrastructure your code passes through.
The honest limits of this test
Three jobs is a small sample, and these were frontend-leaning coding tasks, not large refactors or multi-day migrations. Token counts also depend heavily on how each tool structures its agent loop, so a different task mix could shift the ratio. And Claude Code's count came down from an earlier 360,000 per job after the testers removed their own settings files from the setup, which is a reminder that your configuration also shapes consumption, not just the tool.
A decision framework for picking your AI coding tool in October 2026
Based on the same-model test data and the subscription landscape this month, here is the checklist I would use:
- If quality is your top concern: stop comparing tools and compare models. Sonnet 5.5 passed 186 of 186 in both tools. Model choice drives correctness.
- If you pay per token: tool design now matters for cost, not quality. Cursor used 2.9x fewer tokens in this test, which compounds at API scale.
- If you already have a Claude subscription: Claude Code is included and tied on quality. There is no reason to add a second tool unless you specifically want an editor UI.
- If you want the cheapest usable setup: Cursor's free tier with Auto works, but expect roughly double the latency and a few failed checks versus paid Sonnet 5.5.
- If you are drawn to the X Subscription bundle: value it as a Grok-plus-editor bundle. If you want Claude inside Cursor, budget for API-priced add-ons from the Premium tier up.
- If you are locked into one tool already: nothing in this test says switching improves output. Same model, same results. Switch for cost or workflow fit, not benchmark scores.
The uncomfortable summary: the editor wars are mostly over. The model is the product now, and the tool is the packaging. What you are actually shopping for in 2026 is the packaging that fits your hands at the price you can accept.
I write about AI developer tools, Java, and backend engineering every week. Subscribe if you want more breakdowns like this, it is free.
Do you use Cursor, Claude Code, or something else daily? Did the same-model result match your experience, or does your workflow tell a different story?
Top comments (1)
Really interesting breakdown — the 2.9x token gap is the part that stuck with me, especially that most of it comes from repeated context being re-read into the model rather than extra reasoning. It matches what I've seen in practice, where identical pass rates hide very different cost profiles. One question: given the caveat that Cursor didn't report cache-written tokens, do you think accounting for those would meaningfully close the gap, or is the repeated-context behavior still the dominant cost?