Gemini 4 Carbon vs Fable 5.2, Opus 5.5, GPT 6 Astra, and Grok 4.7 (November 2026)

Model	Input / output per 1M tokens	Context	AA Intelligence Index
Claude Fable 5.2	$8 / $40	1M	65 (max effort), 62 (high)
Gemini 4 Carbon	$2 / $10 through November 30, then $4 / $20	1M	60 (high effort)
Claude Opus 5.5	$4 / $20	1M	59 (max effort)
GPT 6 Astra	$10 / $50	1.05M	57 (max effort)
Grok 4.7	$2 / $6 ($4 / $12 from 200K tokens)	500K	46

Google announced Gemini 4 Carbon on October 28, 2026, positioning it as a model for cyber defenders with a gated rollout through its Fairwind program. On November 10 it became generally available to paid Gemini API developers, Google AI Ultra subscribers, and Gemini for business, one day after Anthropic launched Claude Fable 5.2. The result is the first moment when Google has a credible entry in every price band of the frontier. This post compares Carbon against each of the other four, with numbers current as of November 10.

At announcement Carbon ranked first on Artificial Analysis Intelligence Index v5.0 with a score of 59. Artificial Analysis has since regraded models, and Carbon at high effort now scores 60, behind Fable 5.2.

Access and pricing

Carbon is proprietary and hosted, with no downloadable weights. It takes text, images, speech, and video as input and outputs text, with a 1M token context window. Google AI Ultra subscribers get it in the Gemini app, AI Mode in Search, and Antigravity, at $99.99 per month (Ultra 5x) or $199.99 per month (Ultra 20x). Vertex AI availability and regions are unspecified, and Google has not published a country by country rollout.

On the API, the introductory price is $2 input and $10 output per million tokens through November 30, 2026. After that it is $4 and $20, identical to Opus 5.5. Google has announced a 95% discount on cached input, which implies roughly $0.10 during the promotion and $0.20 after, though those are derived figures, not published rates. Budget against the standard rate, since that is what December looks like.

A less restricted cyber model remains available only through the gated Fairwind program. The generally available model ships with guardrails, and Google's cybersecurity benchmark results were run without them.

Carbon vs Fable 5.2

Benchmark	Carbon	Fable 5.2
AA Intelligence Index	60 (high)	65 (max), 62 (high)
FrontierSWE v2	61.2%	69.9%
FrontierCode 2.0	40.6%	49.1%
Terminal Bench Science 0.1	64.8%	74.6%
Terminal Bench 4.0	65.2%	61.6%
AutomationBench	53.9% at $2.12 per task	51.4% at $4.64 per task
Cost per index task	$7.76 (high), $6.59 (medium, score 58)	$9.08 (high), $26.54 (max), $3.80 (medium, score 56)

Fable is the stronger model on hard software engineering and science, leading by 8.7 points on FrontierSWE v2, 8.5 on FrontierCode 2.0, and nearly 10 on Terminal Bench Science. Carbon takes AutomationBench and does it at less than half the cost per task. It also leads on Terminal Bench 4.0, though that row compares Artificial Analysis's runs of each model's best listed configuration, and Carbon's lead there is small enough to treat as a tie.

On the cost curve, the two cross over. At medium effort Carbon scores 58 for $6.59, while Fable scores 56 for $3.80. At high effort Carbon scores 60 for $7.76 and Fable scores 62 for $9.08. Fable at max effort is the only configuration that reaches 65, and it costs more than three times Carbon at high. Choose Fable when you need the top of the range, and Carbon when you need most of it at a lower and more predictable bill.

Carbon vs Opus 5.5

This is the closest matchup. One index point separates them (60 against 59), and the benchmark rows split: Opus leads FrontierCode 2.0 (45.5% against 40.6%) and edges FrontierSWE v2 (62.3% against 61.2%), while Carbon leads AutomationBench (53.9% against 42.5%) and Terminal Bench Science (64.8% against 58.7%). Through November 30 Carbon costs half as much. After that the list prices are identical, so the choice shifts to ecosystem, tooling, and how each performs on your own tasks. Opus has the cheaper measured cost per task on some third party runs, and Carbon's caching discount has not been fully published, so run both before committing.

Carbon vs GPT 6 Astra

Carbon scores three index points higher while costing a fifth as much per token during the promotion and two fifths as much afterward. Astra leads on FrontierSWE v2 (65.5% against 61.2%) and Terminal Bench Science (68.1% against 64.8%). Carbon leads on AutomationBench by more than 12 points. Astra's cost per index task at max effort is $6.24 against Carbon's $7.76 at high, so on per task spend they are close despite the very different rate cards. Astra keeps its advantages in zero data retention support and in long context recall, where OpenAI reports 96.3% on the 8 needle MRCR test. For the full Astra picture see Fable 5.2 vs GPT 6 Astra.

Carbon vs Grok 4.7

Carbon has the same $2 input price as Grok 4.7 through November 30, 14 index points more (60 against 46), and a Terminal Bench 4.0 score of 65.2% against Grok's 37.6%. Carbon also has twice the context window. Grok retains a real advantage on output price ($6 against $10, then $20 after the promotion) and on speed, which matters for output heavy, low difficulty work. Because Grok's measured cost per task has run higher than its rate card suggests, as we showed in Grok 4.7 vs Opus 5.5, Carbon looks like the better frontier option at this price point on any task where quality matters.

Where Carbon is genuinely different

Security is Carbon's origin story, and its results there are the strongest in the field on published numbers: 88.2% on SEC bench Pro (finding vulnerabilities), 54.8% on ExploitGym (demonstrating exploitability), and 71.0% on CWE bench (patching). Anthropic reports that Mythos 5.2 rediscovered 89.6% of known vulnerabilities in its internal test but gives no Carbon results, so there is no clean cross vendor comparison yet. We unpacked these in our Carbon security benchmark analysis. The practical catch is that the generally available model is guarded, and the unrestricted version is gated. Fable 5.2 blocks offensive security work too, so for teams doing security engineering, Carbon through Fairwind is currently the more direct path.

Google also reports safety improvements: successful cheating on Terminal Bench 4.0 reduced to 0.0%, high severity misalignment alerts in internal coding environments down almost 80%, and a prompt injection attack success rate of 0.0% on the Gray Swan benchmark. These are Google's own figures.

What is not yet verified

Several of Google's launch claims arrived without published scores. Google says Carbon is first on CursorBench 4.0 and WANDR at under half the next best model's cost, but the public CursorBench table shows Fable 5.2 at 60.0% and $32.79 per task with no Carbon row. Wiz's reported discovery of previously unknown vulnerabilities in under two hours is not independently documented. Carbon is also less than a day into general availability, so independent runs are thin. Treat the index scores and cost per task from Artificial Analysis as the most reliable data, and the rest as claims to test.

How to route

If your workload is	Lean toward	Why
Hard engineering, science, long horizon autonomy	Fable 5.2	Leads Carbon by 8 to 10 points on those rows
Business workflow automation	Carbon	Best AutomationBench score at under half Fable's cost
General agent work at the lowest frontier price	Carbon, through November 30	$2 / $10 for a model scoring 60
Everyday coding with a stable cached prefix	Opus 5.5	Leads on FrontierCode, strong cost per task
Defensive security engineering	Carbon via Fairwind	Strongest published vulnerability results
Simple high volume work	Grok 4.7 at low effort	Cheapest output tokens
Math, long context recall, zero retention	GPT 6 Astra	Strengths sit outside the published comparisons

FAQ

Is Gemini 4 Carbon generally available? Yes, as of November 10, 2026, for paid Gemini API developers, Google AI Ultra subscribers, and Gemini for business. The less restricted cyber version remains gated through the Fairwind program.

How much does Gemini 4 Carbon cost? $2 input and $10 output per million tokens through November 30, 2026, then $4 and $20. Cached input is 95% off the input price.

Does Carbon beat Fable 5.2? Not on capability: Fable scores 65 on the Artificial Analysis index at max effort against Carbon's 60, and leads by 8 to 10 points on hard engineering and science benchmarks. Carbon leads on business workflow automation and is cheaper per task on most configurations.

Is Carbon better than Opus 5.5? They are within one index point of each other. Carbon is half the price until November 30, after which the list prices match.

What is the context window of Gemini 4 Carbon? 1M tokens, per Artificial Analysis. Google's announcement did not state it, and a maximum output limit has not been published.

When this applies to your stack

Five models now sit within about 15 index points of each other at prices spanning a factor of five, and the ranking changes depending on whether you measure by capability, rate card, or cost per completed task. The only way to pick well is to run your own tasks through the two or three candidates that fit your constraints and measure the result. Building that evaluation harness and the routing layer behind it is the kind of applied AI work we do at Contra Collective.