Gemini 4 Carbon pricing doubles 1 Dec 2026: $4/$20 vs Fable 5.2 and GPT-6.1 Sol

Summary. Google opened Gemini 4 Carbon (gemini-4-carbon) to all paying Gemini API developers and Gemini for business customers on 10 November 2026 at an introductory $2 per million input tokens and $10 per million output tokens, the same as Gemini 4 Argon. From 1 December 2026 the price is $4 and $20, cached input rises from $0.10 to $0.20, and the Batch, Flex and Priority rates double on the same day. Caching is implicit only, thinking tokens bill as output, and Carbon is not on the free tier. For comparison, Anthropic released Claude Fable 5.2 on 9 November at $8 and $40, while OpenAI's GPT-6.1 Sol stays at $2 and $10 for prompts up to 272,000 tokens. From December, Carbon's per-token rate is half Fable 5.2's and double GPT-6.1 Sol's. On Vals AI's independent Terminal-Bench 4.0 run, updated 9 November, Fable 5.2 scores 65.15% at $14.56 per task, Carbon 63.13% at $7.45 (already priced at $4/$20) and GPT-6.1 Sol 55.05% at $1.72. Cursor's own CursorBench 4.0 table now ranks Carbon first. Re-baseline budgets at $4/$20, finish one-off batch work by 30 November, and send only hard tasks to Carbon.

About this article. This guide is for engineering leads, FinOps owners and CTOs who run or are evaluating Gemini, Claude or OpenAI models in production. We checked every price, date, model ID and benchmark figure on 11 November 2026 against Google's launch posts, pricing page and developer guide, Anthropic's and OpenAI's pricing, model and deprecation pages, OpenAI's changelog, and the Vals AI and Cursor benchmark pages. eCorpIT did not run its own benchmarks. Vendors change prices often, so re-check before you commit; no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What changes on 1 December 2026

Google's launch post, dated 10 November 2026, puts the change in a footnote: once the introductory period ends on 30 November 2026, Carbon costs $4 per million input tokens and $20 per million output tokens. The footnote covers only the publicly available Carbon model. Carbon for cyber, which vetted security teams reach through Google's Fairwind Program, is a separate offering; Google began giving it to a subset of cyber defenders in the week of 28 October.

The Gemini API pricing page now carries a full Carbon section, with both dates on every paid tier.

Gemini 4 Argon has exactly the same schedule on the same page, so any Google workload on either model doubles on the same day.

Two details are easy to miss. Grounding with Google Search is billed per query on top of tokens: 5,000 free search requests a month shared across Gemini 3 and newer models, then $14 per 1,000, with a cap of 1,500 requests a day. And neither the pricing page nor the developer guide says which time zone the switch uses or how a batch job that straddles midnight on 30 November is billed.

More Google increases follow. Gemini 3.8 Flash and Gemini 3.6 Flash double on 1 January 2027, with 3.8 Flash moving from $0.75 and $3.75 to $1.50 and $7.50, as our guide to the Gemini 3.8 Flash price increase explains.

Carbon vs Argon: same price, different appetite

Google prices the two models identically but aims them at different jobs. Its launch post keeps Argon as the default for most software engineering and knowledge work, and positions Carbon for complex, multi-step coding and debugging and open-ended research, where it spends more tokens than Argon to explore, reason and verify. The developer guide adds that Carbon uses more thinking tokens per turn, with higher latency.

Independent data shows the token story depends on the workload. On Vals AI's Terminal-Bench 4.0 leaderboard, Carbon used 12.46 million tokens per task against Argon's 16.66 million, cost $7.45 per task against $17.64, and scored 63.13% against 57.58%. Long agentic tasks that Argon retries or loops on may finish sooner on Carbon. Short interactive turns, where Carbon thinks longer before answering, will cost more. Measure both on your own tasks before you assume either direction.

Moving a workload from Argon or Gemini 3 to Carbon also changes request handling. The developer guide lists these points:

• Thinking level is the main cost control. Carbon supports low, medium (default) and high, all at the same per-token price. Google recommends evaluating at all three and keeping the lowest level that passes.

• Explicit caching is not supported. Remove cachedContents and caches.create calls. Implicit caching applies automatically to prompts of at least 4,096 tokens.

• Sampling parameters are ignored. temperature, top_p and top_k have no effect.

• Output limits need headroom. Google suggests a max_output_tokens of at least 32,768 because thinking draws on the same budget. Synchronous responses stop at 65,536 tokens; longer output needs background execution or client-side continuation.

• Timeouts. Hard prompts can take several minutes; Google recommends background execution or HTTP client timeouts of at least 15 minutes.

Gemini 4 Carbon vs Claude Fable 5.2 vs GPT-6.1 Sol: the price sheet

All three target hard, long-running work and accept about a million tokens of context, but they charge for it differently. The table draws on Google's developer guide and pricing page, Anthropic's pricing and Fable 5.2 pages, and OpenAI's GPT-6.1 Sol, pricing, changelog and GPT-6 guide pages.

Fable 5.2 is the expensive option per token, though Anthropic priced it 20% below Fable 5.1, which lists at $10 and $50 with $0.25 cache reads. Anthropic's models overview tells most users to start with Claude Opus 5.5 ($4 and $20) and to move to Fable 5.2 for demanding reasoning and long-horizon agent work, or when Opus 5.5 at higher effort still fails their evaluations. That makes Opus 5.5, not Fable, Carbon's like-for-like price rival from December. Our best LLM for coding comparison covers Opus 5.5, Argon and GPT-6.1 Sol in depth, and our Claude Fable 5.2 pricing and API changes guide covers the Anthropic side.

Benchmarks: independent results first, vendor claims labelled

eCorpIT did not run its own tests for this article. The figures below come from Vals AI, which runs every model through the same harness, and from vendor announcements, labelled as such.

Cost per solved task is the cost per task divided by the score; for Carbon, $7.45 / 0.6313 = $11.80.

Three caveats. First, Vals runs Terminal-Bench 4.0 in mini-swe-agent with bash as the only tool, reports the mean of three full runs, and uses 66 long tasks whose median carries a four-hour expert estimate. Your coding agent, IDE or tool set will produce different numbers. Second, Vals lists Carbon's and Argon's token costs as $4 and $20, so these per-task figures already reflect the December price; at the introductory rate the same Carbon run would cost about $3.73 per task ($7.45 / 2). Third, Claude Opus 5.5 ties Fable 5.2 at 65.15% on the same leaderboard for $12.04 per task, which is another reason to test Opus before Fable.

Vendor claims and Cursor's table. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at less than half the cost per task of the next best external model, and second on the Multimodal Knowledge Agent benchmark at under a sixth of the leader's average cost per task. Cursor's own CursorBench 4.0 table, as fetched on 11 November, lists Carbon at its high setting first with 60.4% at $9.22 per task, just ahead of Fable 5.2 at max effort with 60.0% at $32.79; Cursor notes that small score gaps may not be statistically meaningful. Cursor prices each model at its published per-token rates but does not say which of Carbon's two prices it applied. If it used the introductory rate, doubling Carbon's cost gives about $18.44 per task, still well below Fable 5.2's $32.79 but a gap of about 1.8 times rather than 3.6 times. GPT-6.1 Sol is not on Cursor's table yet, and we did not check WANDR or AutomationBench. OpenAI's description of GPT-6.1 Sol, as close to GPT-6 Astra on complex work at a lower cost, is also a vendor claim.

Worked cost example: one team, two views

Count your real prompt sizes before you rely on these numbers; the three vendors' tokenizers produce different counts for the same text.

View 1: identical token profile. An analysis and coding agent uses 1,000 million input tokens a month, 85% served from cache (850 million) and 150 million uncached, plus 80 million output tokens including thinking. For Fable 5.2 and GPT-6.1 Sol we show a range: uncached tokens billed at the base input rate (low) or as five-minute or 30-minute cache writes (high). Rates come from the three pricing pages.

At equal token counts, a November pilot on Carbon costs the same as GPT-6.1 Sol, and the December invoice doubles. Fable 5.2 costs about twice Carbon's December rate (1.9 to 2.1 times). Gemini 4 Argon would also show $2,370 here, since it shares Carbon's price.

View 2: per task, using Vals' measured token habits. The equal-token view hides how differently each model spends tokens on the same job. Take 500 long agent tasks a month at Vals' Terminal-Bench 4.0 costs and solve rates.

On this benchmark Carbon buys about 40 more solved tasks than GPT-6.1 Sol for $2,865 more, roughly $71 per extra solved task ($2,865 / 40.4). Whether that is worth paying depends on what a failed task costs you in engineer time. Against Fable 5.2, Carbon solves about ten fewer tasks for $3,555 less.

What to do before and after 30 November: caching, batch and routing

1. Re-baseline every budget at $4/$20. Any pilot run between 10 and 30 November shows half the December cost. Put the December rate in your forecast, your gateway's price table and your cost alerts now, and tell finance that the first December invoice doubles for the same usage.

2. Finish one-off batch work by 30 November. Evaluation sweeps, document backfills and similar jobs planned for December are cheaper now: Carbon's Batch rate moves from $1 and $5 to $2 and $10. Because Google does not say how a job that straddles the switch is billed, aim to complete jobs, not just submit them, before the date.

3. Make implicit caching work for you. Carbon has no explicit cache and no write fee. Cached input costs $0.20 from December, a 95% discount, but only for prompts of at least 4,096 tokens that repeat a prefix the system has already seen. Put the stable parts first (system prompt, tool declarations, repository map, policy documents) and the changing user turn last. Track the cached-token count in each response's usage data; if the hit rate is low, restructure prompts before you change models. Fable 5.2 reads cache at 2.5% of input but charges $10 or $16 per million to write it, and GPT-6.1 Sol reads at 5% and writes at $2.50 or $5. On workloads with poor cache reuse, Carbon's lack of a write fee narrows the gap.

4. Use thinking level and output caps as the dial. Carbon bills thinking tokens as output at $20 per million from December. Google's guide suggests low for simpler and cost-sensitive work, including light analysis and writing, medium for most tasks including complex code, and high for the hardest reasoning. Set max_output_tokens per route so one runaway task cannot burn through the 1M window.

5. Route instead of picking one model. A gateway that sends routine work to a cheaper model and escalates hard tasks to Carbon or Fable often beats any single choice. In View 2, sending 70% of tasks to GPT-6.1 Sol and 30% to Carbon costs 350 × $1.72 + 150 × $7.45 = $1,719.50 a month, 54% less than all-Carbon. Route by task class (repository size, test count, prior failure), not at random, and log which route solved each task. Our hybrid routing decision framework covers rule design. If OpenAI is one of your routes, first check that no calls still use gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 or o3-pro, which OpenAI removes from the API on 11 December 2026; our GPT-5 and o3 shutdown migration plan has the steps.

6. Watch the cheap tier too. If you route easy work to Gemini 3.8 Flash, its price doubles on 1 January 2027. Our comparison of Gemini 3.8 Flash, Claude Haiku 5.5 and GPT-6 Luna covers the alternatives for that tier.

Decision table: which model, from 1 December

India, EU and US notes

Google. The pricing page says paid-tier Gemini API content is not used to improve Google's products. The developer guide's examples call the generativelanguage.googleapis.com endpoint, and we could not find an India-region option for Carbon in Google's guide or pricing page. Enterprises that need regional controls, provisioned throughput or volume discounts should ask Google's sales team, which offers those through the Gemini Enterprise Agent Platform, and confirm whether Carbon is available there.

Anthropic. Fable 5.2 runs on the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry and Claude Platform on AWS. US-only inference through inference_geo adds a 1.1x multiplier to every token category on Anthropic-operated platforms. Fable 5.2 requires 30-day data retention and is not offered under zero data retention unless Anthropic expressly authorises it, which matters for workloads with strict retention rules.

OpenAI. Regional processing for GPT-6.1 Sol adds 10%. Its Ultrafast mode supports US and EU data residency, and Fast mode supports EU residency.

For Indian teams processing personal data, record where inference runs and how long each vendor keeps prompts, and reflect both in your privacy notices and vendor records for DPDP Act and GDPR purposes. The vendor settings above are inputs to that work, not a compliance answer, and this article is not legal advice.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI gateway and model routing service helps teams evaluate Gemini, Claude and OpenAI models on their own tasks, route between them and track cost per route before a price change lands. Talk to us at /contact-us/.

Last updated: 11 November 2026.

References

1. Gemini 4 Carbon is now available, 10 November 2026 — Google

1. Gemini Developer API pricing — Google AI Studio

1. What's new in Gemini 4 Carbon — Google AI Studio

1. Claude Fable 5.2 — Claude Platform Docs, Anthropic

1. GPT-6.1 Sol model — OpenAI API

1. Terminal-Bench 4.0 leaderboard and methodology, updated 9 November 2026 — Vals AI

1. CursorBench 4.0 — Cursor

1. Introducing Gemini 4 Carbon, 28 October 2026 — Google

1. Pricing — Claude Platform Docs, Anthropic

1. Pricing — OpenAI API

1. Changelog — OpenAI API

1. Using GPT-6 — OpenAI API

1. Models overview — Claude Platform Docs, Anthropic

1. Gemini 4 Carbon benchmarks, cost and capabilities — Vals AI

1. Claude Fable 5.2 benchmarks, cost and capabilities — Vals AI

1. GPT-6.1 Sol benchmarks, cost and capabilities — Vals AI

1. Gemini 4 Argon benchmarks, cost and capabilities — Vals AI

1. Deprecations — OpenAI API