Claude Opus 5.5 vs Gemini 4 Carbon: A 12% Ceiling and a Third of the Bill

Claude Opus 5.5 and Gemini 4 Carbon are the top two models on SemiAnalysis's new GDPx benchmark, and the gap between them is a little over one point: 12.0% for Opus 5.5, 10.9% for Carbon. The more useful number sits underneath the score. Carbon got there for $16.29 a task; Opus 5.5 needed $51.41. That is the whole decision in one line, with a catch the leaderboard cannot show. Opus 5.5 has been generally available since September 22, 2026. Carbon was announced on October 28 and is still restricted to cyber-defence partners through Google's Fairwind program, with no public model ID. A third of the bill is only a bargain if you can place the order.

GDPx v1.0 is a new benchmark, released this week, and it is the first independent evaluation to score these two models on the same tasks with the same harness and report what each run cost. That timing matters more than the ranking. Six weeks after Opus 5.5 shipped, and days after Carbon was announced, there is finally a measured comparison rather than two vendor tables placed side by side.

What each model is, before the table

Claude Opus 5.5 is Anthropic's current flagship for long-running agentic coding and knowledge work. It replaced Opus 5 with a 20% price cut and a documented 1M-token context window, and it is the model Anthropic positions for Claude Code and the Claude apps.

• Price — $4 per million input tokens, $20 per million output. Batch requests are 50% off at $2 / $10.

• Context and output — 1M input tokens, 128K output on the synchronous API, 300K output on the Batch API with the extended-output beta header.

• Caching — a five-minute cache write costs $5 per million tokens, a one-hour write $8, and a cache read $0.20.

• Thinking — adaptive thinking is always on and cannot be disabled; the default effort level is medium, and the reliable knowledge cutoff is June 2026.

• Where — the Claude API under claude-opus-5-5, Amazon Bedrock, Google Cloud through both the Google-managed and Anthropic-managed paths, Microsoft Foundry, and the Claude Platform on AWS.

Gemini 4 Carbon is Google's newest frontier model, announced by Koray Kavukcuoglu on October 28 as the company's strongest model for cybersecurity, long-horizon agents and complex software engineering. Google's announcement gives it three headline claims and one conspicuous omission: no context window is published. The benchmark claims are Google's own — SEC-bench Pro 88.2% for vulnerability discovery, ExploitGym 54.8% for exploitation, and CWE-bench 71.0% for patching — and should be read as vendor-reported, not independently reproduced.

• Price — $2 per million input tokens and $10 per million output during an introductory period, with cached input at 95% off; the standard rate is later $4 / $20, and Google has not said when the introductory pricing ends.

• Access — Fairwind only, to a subset of cyber defenders and trusted partners, with broader developer, enterprise and consumer access promised "as soon as possible" and no date.

• Endpoint — none public. The Gemini API still exposes gemini-4-argon, not Carbon.

• Context — not stated in the announcement. Long-context retrieval results have circulated in launch-week coverage, but the window itself is not a published specification.

That asymmetry is the first thing to understand about this matchup. One model is a product you can call today from the Claude API, four cloud platforms and a dozen SDKs. The other is a capability demonstration with a waiting list. Everything that follows compares them as models; only one of them is currently a deployment option.

GDPx: 206 tasks, a 12% ceiling and a 3.2x cost gap

SemiAnalysis built GDPx as a successor to OpenAI's GDPval, with 14 data partners and 206 tasks spread across 11 professional domains: finance, legal, biology, materials science, chip design, hardware design, cybersecurity, software engineering, alignment, quant trading and recursive self-improvement. Each task gets an eight-hour default budget, three runs are averaged, and the domains are weighted equally. Only 13 of the 20 models tested appear on the public board.

• Claude Opus 5.5 — 12.0%, first, at $51.41 per task and roughly 548,000 output tokens per task.

• Gemini 4 Carbon — 10.9%, second, at $16.29 per task and roughly 594,000 output tokens per task.

• GPT-6 Astra — 9.6%, third, at $53.99 per task.

• Claude Fable 5.1 — 8.5% at $67.16, the most expensive model on the board.

Two readings compete here, and both are correct. The first is that Carbon delivers about 91% of Opus 5.5's score for about 32% of the cost, which is a genuine value argument and the one 9to5Google led with. The second is that neither model finishes seven of every eight professional tasks it is given. A one-point lead at 12% is not a solved problem; it is the current edge of what agents can do when the work is long, open-ended and judged by a domain expert.

SemiAnalysis published a second view with cybersecurity, biology and alignment removed, and the ordering changes: GPT-6 Astra moves to first, and Carbon falls below GPT-6.1 Sol and Claude Sonnet 5.5. The company says that view better matches its subjective impressions. The mechanism is guardrails — a model that declines a task scores zero for it, and Carbon's safety behaviour is stricter in exactly the domains where its cyber strengths are largest. So Carbon's second place is partly a capability result and partly a permission result. If your work looks like the removed domains, the headline ranking is not describing you.

Where Carbon is actually ahead

On Artificial Analysis's Intelligence Index v5.0, Carbon sits at 59 and Opus 5.5 at 58. One point, on a composite that was rebased on October 27, is not a gap worth migrating a production system for. It does establish that Google's model is at the frontier rather than behind it, which is a stronger claim than the gated rollout makes easy to verify.

The cyber numbers are where Carbon's case is clearest, and they are also where the sourcing is weakest. Google's own harness puts it at 54.8% on ExploitGym, 71.0% on CWE-bench and 88.2% on SEC-bench Pro. Those are large, specific claims about vulnerability discovery and patching, and they are exactly the kind of claim that needs an independent rerun before a security team budgets against it. GDPx's cyber domain is one input to that, not a replication of those three suites.

Price is the third advantage, with an asterisk. At $2 / $10, Carbon's introductory rate is half of Opus 5.5's $4 / $20. At the standard rate of $4 / $20 it is the same. The 95% cached-input discount is real but narrows at list price: $0.10 per million cached tokens at the introductory rate, $0.20 at the standard rate, against Opus 5.5's $0.20 cache read. The cost advantage is largest today, while the discount lasts, and Google has not said how long that is.

Where Opus 5.5 still leads

Opus 5.5's strongest independent result is not the GDPx score but the coding profile. Artificial Analysis's Coding Agent Index has it at 62 against Carbon's 59, and that index is built from agentic coding runs rather than single-shot completions. For a model whose stated purpose is long-running software work, that is the closer measurement.

Anthropic's own launch figures point the same way, and they are vendor-reported: Terminal-Bench 4.0 at 66.4%, FrontierCode v1.1 at 54.4%, CursorBench 4.0 at 57.8% and AutomationBench at 40.0%. The harness matters. Artificial Analysis measured the same model at 59.6% on Terminal-Bench 4.0 at xhigh effort, and Vals AI reported 65.2%. A seven-point spread between the vendor and one independent harness is normal for agentic benchmarks, and it is the reason a single row should never decide a purchase.

The unglamorous advantage is availability. Opus 5.5 has a published context window, a documented output limit, a batch tier, a fast mode at $8 / $40, a migration guide and six weeks of production use behind it. Carbon has a blog post, a partner program and a price list. For a team that needs an answer this quarter, that difference decides more cases than 59 versus 58.

What a task actually costs

GDPx's cost figures are measured token spend under an eight-hour budget, not list prices, so they include how much each model chooses to write. Carbon used about 8% more output tokens per task than Opus 5.5 and still cost a third as much, because the introductory rate is half and the input side is where long agent runs accumulate. On 1,000 tasks, the reported gap is roughly $51,000 against $16,000.

At list price the arithmetic is less dramatic. A request with 1M input tokens and 250K output tokens costs $9 on Opus 5.5 at $4 / $20, $4.50 on Carbon at the introductory $2 / $10, and $9 on Carbon at its standard $4 / $20. Opus 5.5's Batch API at 50% off lands at $2 / $10 — the same as Carbon's introductory rate — for work that can wait. When Carbon's introductory period ends, the rate-card difference disappears entirely; what remains is the per-task difference in token use, which is small.

This is the part of the comparison most likely to be quoted out of context. "A third of the cost" is true of GDPx runs at today's promotional pricing. It is not a permanent property of the model, and it is not the ratio you will see on a batch workload at standard rates.

The availability gap is the real decision

Claude Opus 5.5 is live on OrcaRouter as anthropic/claude-opus-5.5 at the provider's $4 / $20 list price, with no markup added, alongside 200+ models on one key. Routing it is a model-string change, and a weighted slice with automatic failover is the cheap way to test it against your current model without betting a production path on the result.

Gemini 4 Carbon is not on OrcaRouter and cannot be. Google has published no model ID and no public endpoint, so there is nothing to route to; the only honest statement is that access runs through Google's Fairwind program and, eventually, the vendor's own API. When a public endpoint does appear, the pass-through model means the list price is what you pay here the day it changes rather than a repriced version of it.

That leaves a decision that is narrower than the benchmark table suggests:

• Security work with Fairwind access — trial Carbon on your own triage and patching queue. Its vendor cyber numbers are the strongest published, but they are unreplicated, and the GDPx domain-reweighting result says its advantage is domain-dependent.

• Long-running coding agents — Opus 5.5. It leads the independent coding index, has a documented 1M-token window and 128K output, and is available on the platforms most enterprises already contract with.

• Cost-sensitive batch work — Opus 5.5's Batch API at $2 / $10 matches Carbon's introductory rate today, and does not depend on a discount that Google can end without notice.

• A straight capability bet — wait. A one-point Intelligence Index lead and a 1.1-point GDPx lead are inside the noise of harness choice, and neither model is close to finishing the professional tasks GDPx sets.

What would change the answer

Three things are worth watching, and none of them is another vendor benchmark. The first is a public Carbon endpoint. Until Google ships a model ID, every cost comparison is hypothetical for anyone outside Fairwind, and the introductory price is a promise rather than a bill. The second is the end of that introductory period; at $4 / $20, Carbon's rate card is identical to Opus 5.5's, and the value argument collapses back to token efficiency. The third is an independent replication of the cyber results. SEC-bench Pro at 88.2% and ExploitGym at 54.8% are the numbers that would justify a security budget, and they currently rest on one lab's harness.

For now the honest summary is narrower than either vendor would like. Opus 5.5 is the better model to build on, because it is the one you can build on. Carbon is the more interesting model to watch, because it matches the frontier at a lower price and has not yet been allowed to prove it in public. The 12% ceiling belongs to both of them.