Gemini 4 Carbon vs GPT-6.1 Sol: Prices, Benchmarks and Which One Fits Your Job
Gemini 4 Carbon and GPT-6.1 Sol are the two flagship workhorses of Google and OpenAI — and right now they cost exactly the same. Both charge $2 per million input tokens and $10 per million output tokens. That price parity ends on 30 November, when Carbon's introductory rate expires and its price doubles to $4 and $20. On the newest independent maths test, Carbon leads Sol 85.4% to 80.6% — but it costs about fifteen times as much per problem to get there. The two models are not really competing on the same job, and the right choice depends on what you are asking the model to do, not on which one tops a leaderboard.
Carbon is Google's strongest frontier model, announced on 28 October and generally available from 10 November, built for deep reasoning and cybersecurity. Sol is OpenAI's middle model, released on 29 September at DevDay, one week after the original GPT-6 Sol it replaced.
The price table — and the deadline that changes it
At standard rates, for short-context requests, the two models are priced identically today. The table below uses the rates as published on each provider's pricing page, with the expiry that matters most stated plainly.
| Price per 1M tokens (Standard, short context) | Gemini 4 Carbon | GPT-6.1 Sol |
|---|---|---|
| Input | $2.00 (until 30 Nov 2026) | $2.00 |
| Cached input | Not stated in this comparison | $0.10 |
| Output | $10.00 (until 30 Nov 2026) | $10.00 |
| From 1 Dec 2026 | $4.00 / $20.00 | $2.00 / $10.00 (unchanged) |
| Free tier | None | None (API is pay-as-you-go) |
| Context | 1M tokens | Short-context rates apply up to 272K input tokens |
Two details deserve attention. First, Carbon's $2/$10 is introductory and expires on 30 November — quote it without the expiry and the comparison misleads anyone reading after that date, because from 1 December Carbon costs twice as much as Sol on every token. Second, Sol's pricing has more structure: cached input costs $0.10 per million tokens, just 5% of the fresh-input rate, and cache writes cost $2.50; Batch and Flex processing each halve the standard rates for work that can wait or tolerate capacity errors. Above 272,000 input tokens, Sol moves to long-context rates of $4 input and $15 output for the whole request. Carbon's cached-input and batch rates are not stated here because they were not re-verified in this session.
Until 30 November, token price cannot decide between them. From 1 December, Sol is half Carbon's price, and its caching and batch discounts widen the gap further. Anyone budgeting past November should price Carbon at $4/$20, not at the launch rate.
![]() |
AI-generated illustration (generated 14 Nov 2026)
Benchmarks: Carbon wins the maths test, Sol wins the value test
The freshest independent comparison comes from MathArena, the ETH Zurich benchmark that tests models on newly released competition and research maths problems. On its overall table, read on 14 November, Carbon at "high" reasoning scores 85.4% expected performance, against Sol at "max" reasoning with 80.6%. The gap is 4.8 points, and it is the first time in a long while that a non-OpenAI model has led that board. But MathArena also publishes the cost of one run on one problem: $15.70 for Carbon against $1.07 for Sol. Carbon's lead costs about fifteen times as much per problem, and part of its score is statistically estimated rather than measured, because cost meant Carbon was run only on ArXivLean and will be added to LeanLean later.
On OpenAI's own published charts, Sol's strengths are agentic coding and professional work rather than pure maths. At high reasoning effort it scores 75.2% on DeepSWE, a test of long software-engineering tasks in real codebases, at $0.65 a task — ahead of GPT-6 Astra's best of 74.1% at $4.43 a task. It reaches 71.4% on OSWorld 2.0 offline computer-use tasks at maximum effort for $1.27 a task, and 36.1% on AutomationBench business workflows at $0.30 a task. On Terminal-Bench Science it scores 57.0% at maximum effort, and its factual error rate on deliberately hard prompts falls to 4.1% at extra-high effort. These are vendor-reported figures, not independent tests — but the cost-per-task figures alongside them are the useful part, because they show Sol beating or nearly matching a model five times its price.
The closest shared yardstick between the two is the Artificial Analysis Intelligence Index, a composite of ten benchmarks. Carbon scores 59; Sol scores 52 at maximum reasoning. Treat that as a rough indication, not a head-to-head: the two scores come from different benchmark sets and settings, and no independent test has yet run both models on the same suite under the same conditions. The honest summary is that Carbon looks stronger on the hardest reasoning, Sol looks stronger on value, and nobody has yet published the test that would settle the rest.
Which one fits which job
![]() |
Choose Carbon when the job is deep reasoning where the answer matters more than the bill: hard mathematics, research-level problems, and the cybersecurity work Google built it for. Its 1M-token context window suits long documents and large codebases held in a single session. The cost is the trade-off — and from December, that trade-off doubles.
Choose Sol when the job is everyday agentic work: coding agents, computer-use workflows, business automation, and professional document work. It is OpenAI's default pick for coding agents for a reason — its DeepSWE score beats Astra's at a seventh of the cost per task, and its caching and batch rates make high-volume use far cheaper than Carbon after November. Sol runs in ChatGPT Work and Codex for paid plans. Its five reasoning efforts — low, medium, high, xhigh and max — let you pay for depth only when the task needs it; low effort still scores 64.4% on DeepSWE at $0.17 a task.
Neither model is a general winner. Carbon's cyber benchmarks were evaluated without guardrails, according to Google's own note, so they describe capability rather than the behaviour of the guarded product. Sol's headline numbers are OpenAI's own. A buyer who picks on a single leaderboard will regret it; a buyer who matches the model to the job — Carbon for the hardest reasoning, Sol for the daily agentic workload — will get the value out of either.
Background: Gemini 4 Carbon is covered in full in our launch article, and our flagship AI model comparison puts Carbon against Claude Opus 5.5 and GPT-6 Astra on the Artificial Analysis index. OpenAI's cheaper tier is GPT-6.1 Terra, at half Sol's price, and the GPT-6 Astra benchmark scores explain why one model's numbers can look so different from one test to another.
![]() |
| AI-generated illustration (generated 14 Nov 2026) |
Key Takeaways
- Carbon and Sol both cost $2/$10 per million tokens today — but Carbon's rate is introductory and doubles to $4/$20 on 1 December, while Sol stays at $2/$10 with $0.10 cached input and Batch/Flex at half price.
- On MathArena, Carbon leads Sol 85.4% to 80.6% — the first non-OpenAI model to top that board in a long while — but costs $15.70 per problem against Sol's $1.07, about 15 times as much.
- On OpenAI's own charts, Sol scores 75.2% on DeepSWE at $0.65 a task, ahead of Astra's 74.1% at $4.43, and 71.4% on OSWorld 2.0 offline at $1.27 a task.
- The Artificial Analysis Intelligence Index puts Carbon at 59 and Sol at 52, but the scores come from different benchmark sets — no independent head-to-head exists, so no winner is declared here.
- Carbon fits deep reasoning, hard maths and cybersecurity with a 1M-token context; Sol fits everyday agentic coding, computer use and business workflows at a lower cost per task.
Conclusion
Gemini 4 Carbon vs GPT-6.1 Sol is a comparison with a deadline built into it. Until 30 November, the two models cost the same per token, and the choice is about capability: Carbon's deeper reasoning against Sol's agentic value. From 1 December, Carbon costs twice as much, and the maths changes with it — Sol becomes the cheaper model on every token, with caching and batch discounts on top. Carbon's MathArena lead is genuine, but it was bought at fifteen times the cost per problem, and no independent test has yet run both models head to head. Pick Carbon for the hardest reasoning you can afford; pick Sol for the daily work you actually do. And if you are budgeting past November, price Carbon at $4/$20 — the launch rate is already expiring.
Sources
- MathArena — overall leaderboard (Carbon 85.4% at $15.70/problem; GPT-6.1 Sol 80.6% at $1.07/problem; read 14 Nov 2026)
- Requesty — OpenAI API pricing (GPT-6.1 Sol $2/$0.10/$10 per 1M tokens, short context; Batch and Flex at half price; as of 3 Nov 2026)
- AI Catchup — GPT-6.1 Sol: Pricing, Benchmarks, Effort Costs, vs Astra and Opus 5.5 (OpenAI's own benchmark charts: DeepSWE 75.2%, OSWorld 71.4%)
- ai-oriens — MathArena report (Carbon GA 10 Nov, Fable 5.2 released 9 Nov; Pichai and Hassabis reposts)



Comments
Post a Comment