Gemini 4 Carbon vs Claude Opus 5.5 vs GPT-6.1 Sol: price, context, access and benchmarks (October 2026)

Gemini 4 Carbon vs Opus 5.5 is the comparison many developers started making on October 28, 2026, when Google announced Carbon as its strongest Gemini model. This explainer adds OpenAI's GPT-6.1 Sol, which OpenAI prices at one-fifth of its flagship, GPT-6 Astra. The short answer: on Artificial Analysis's independent Intelligence Index, they're within four points of each other, but only two of them can be bought. Carbon is limited to vetted cyber defenders for now, so its price and limits are what Google has announced, not what you can use. For the Carbon launch itself, see our Gemini 4 Carbon news story.

This explainer is based on Google's October 28 Carbon post, its Gemini benchmark table and methodology, Anthropic's and OpenAI's pricing pages, model pages and plan pages, OpenAI's Codex pricing page and developer-forum launch post, and results published by Artificial Analysis and Vals AI, all checked October 29, 2026.

The short version

• Price: Carbon will launch at $2/$10 per million input/output tokens, then $4/$20, Google says. Opus 5.5 is $4/$20. GPT-6.1 Sol is $2/$10, rising to $4/$15 for prompts over 272K tokens.
• Context: Opus 5.5 has a 1M-token window and GPT-6.1 Sol 1,050,000 tokens, each with up to 128,000 output tokens. Google hasn't published Carbon's; Artificial Analysis lists 1.0M.
• Access: Carbon: a subset of cyber defenders and Fairwind partners only. Opus 5.5: Claude API, AWS, Google Cloud, Azure, and Claude Pro and above. GPT-6.1 Sol: OpenAI API, and ChatGPT Work and Codex on Plus and above.
• Independent score: Artificial Analysis's Intelligence Index v5.0 puts Carbon (high) at 59, Opus 5.5 (max) at 58 and GPT-6.1 Sol (max) at 55.
• Cost per task on that index: $7.76 for Carbon, $10.04 for Opus 5.5 and $1.27 for GPT-6.1 Sol, per Artificial Analysis.

How much do Gemini 4 Carbon, Claude Opus 5.5 and GPT-6.1 Sol cost?

All prices are US dollars per million tokens, standard tier.

Last verified: October 29, 2026

What changes the bill:

• Carbon's terms are announced, not listed. Google says Carbon "will launch at the same introductory price as Argon," with cached input "priced at 95% off input token price," and that "$4 per 1M input tokens and $20 per 1M output tokens will apply" after the introductory period. Carbon isn't on the Gemini API pricing page, and Google hasn't said when the introductory period ends. Argon's ends November 30, 2026, per that page.
• Opus 5.5 has no long-context surcharge. Anthropic's pricing page says a 900K-token request is billed at the same per-token rate as a 9K-token one. Cache writes cost $5 (five-minute) or $8 (one-hour). Fast mode is $8/$40.
• GPT-6.1 Sol doubles input above 272K. OpenAI's model page says requests over 272K input tokens cost 2x the input and cache rates and 1.5x the output rate for the whole request. Fast mode is $4/$20 and Ultrafast $12/$60, as our GPT-6.1 Sol pricing story explains.
• Thinking tokens count as output on all three vendors' reasoning models.

Worked example (illustrative, not a measurement). One request with a 400,000-token prompt and 8,000 output tokens, at list prices. Google hasn't published a long-prompt rate for Carbon, so none is applied:

Last verified: October 29, 2026

List prices aren't the whole cost, because models spend different numbers of tokens. Artificial Analysis's cost-per-task figures, below, capture that for its own test set.

Context windows, limits and where you can use each

Last verified: October 29, 2026

Google says Carbon goes to "a subset of cyber defenders" starting the week of October 28, then to "additional trusted partners through our Fairwind Program," and later to "developers, businesses, and consumers via our product surfaces as soon as possible." For context, Argon went from Fairwind-only to general availability in 13 days; see our Argon price and plans story.

OpenAI's Codex pricing page says: "In ChatGPT, GPT-6.1 Sol is available in Work and Codex, not Chat." On Enterprise and Edu it is "off by default until an administrator enables it." Our GPT-6 Astra vs 6.1 Sol vs Luna explainer covers which ChatGPT tab runs which model. Anthropic's pricing page marks Opus as included on Pro, Max, Team and Enterprise; our Claude lineup guide covers when to pick Opus over Sonnet or Fable.

Which is best on benchmarks: Carbon, Opus 5.5 or GPT-6.1 Sol?

No single evaluation answers that. Two independent indexes, from Artificial Analysis and Vals AI, have published scores for all three, and each vendor's own table favors its model. The results below are kept apart by who ran them.

Artificial Analysis Intelligence Index v5.0 (independent)

Artificial Analysis launched version 5.0 of its index on October 27, 2026 and added Carbon on October 28. The index is a weighted average of nine evaluations in four equal categories: knowledge work, coding, science and general knowledge. Our explainer on what changed in Intelligence Index v5 covers the method. Scores from version 4.3.2 and earlier aren't comparable.

Last verified: October 29, 2026

Limits to keep in mind:

• Different settings. Carbon ran at high reasoning; Opus 5.5 and GPT-6.1 Sol at max. Artificial Analysis labels the Opus 5.5 run "with fallback." Anthropic's own launch page says that when its production safeguards intervene, some tasks are completed by older Claude models.
• Carbon isn't public. Artificial Analysis says Carbon "is being rolled out to selected users and is not yet publicly available," and doesn't say how it got access. Its public leaderboard, checked October 29, now lists Carbon first at 59, ahead of Opus 5.5 (max) at 58.
• Cost assumptions. Artificial Analysis priced Carbon at the later $4/$20, not the launch price.
• What it measures. Artificial Analysis calls the index "a primarily text-based, English-language evaluation suite."

Per-evaluation results Artificial Analysis has published add detail. On its AA-Omniscience knowledge test, Carbon scores 54 against 48 for Opus 5.5, and Carbon's hallucination rate is 11% against 50% for Opus 5.5 (max). On CritPt v2, a physics research test, GPT-6.1 Sol (max) scores 69% and Opus 5.5 61%. On Terminal-Bench-Science, Opus 5.5 scores 59% and GPT-6.1 Sol 58%; Artificial Analysis says Carbon leads that test but didn't give its score in text. On its separate Coding Agent Index, Carbon (high) running in Google's Antigravity CLI scores 59 against 62 for Claude Code running Opus 5.5 (max). The firm calls GPT-6.1 Sol "within 4 points of Claude Opus 5.5 (max) at about an eighth of the cost per task."

Vals AI (independent)

Vals AI's Vals Index v2.2, checked October 29, 2026, scores Carbon at 70.36% ($10.51 per test), Opus 5.5 at 70.18% ($28.12 per test) and GPT-6.1 Sol at 63.94% ($3.61 per test). Carbon is second and Opus 5.5 third, behind Claude Sonnet 5.5 at 70.85%. That matches the 70.4% Google's table shows for Carbon, which Google says comes from Vals AI.

Google's table (vendor-reported)

Google's Gemini page compares Carbon with Opus 5.5, GPT-6 Astra and others, but not GPT-6.1 Sol. Google ran Carbon at its highest thinking setting and took most rivals' scores from those rivals' own reports.

Last verified: October 29, 2026

Two caveats from Google's methodology: Carbon's OSWorld score is the best of three runs, and GraphWalks was computed by Google. There's also a discrepancy: Google lists Carbon at 65.2% on Terminal-Bench 4.0 "sourced from Artificial Analysis," while Artificial Analysis's own Carbon writeup gives 64%. Anthropic's launch table reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 at xhigh effort, run with its production safeguards on. OpenAI's launch post for GPT-6.1 Sol reported 75.2% on DeepSWE v1.1 at high effort and 71.4% on OSWorld 2.0's offline set at max effort. Those are different harnesses and versions, so they can't be lined up with Google's numbers.

Which should you use?

Last verified: October 29, 2026

If you need a Google model now, Argon is on sale at $2/$10 through November 30; our Argon vs Opus 5.5 vs GPT-6 comparison covers it. For consumer plans across all three companies, see Claude vs ChatGPT vs Gemini.

What's confirmed, reported and uncertain

• Confirmed (official pages): Opus 5.5 and GPT-6.1 Sol prices, limits and plan access; GPT-6.1 Sol's 272K surcharge; Opus 5.5's flat long-context pricing; Carbon's announced price and limited rollout.
• Vendor-reported: every number in Google's table, Anthropic's Terminal-Bench 4.0 result and OpenAI's launch scores.
• Independent, with limits: Artificial Analysis's Intelligence Index v5.0 scores and costs, at different effort settings; Vals AI's Vals Index v2.2 scores and costs.
• Uncertain: Carbon's release date, model ID, context window, output limit, long-context pricing and batch pricing; when its introductory price ends; the 64% vs 65.2% Terminal-Bench 4.0 discrepancy.