OpenAI shipped GPT-5.5 on April 23, 2026 — seven days after Anthropic shipped Claude Opus 4.7. Both arrived claiming coding leadership. OpenAI’s model page lists GPT-5.5 at $5 input and $30 output per million tokens, double GPT-5.4’s $2.50 and $15. Anthropic held Opus 4.7 flat at $5 and $25.

October 2026 update: both models have been superseded. Claude Opus 5.5 launched on September 22 at $4/$20, which makes Opus 4.7 (now listed as legacy) more expensive than its successor. For the current matchup, read Claude Opus 5.5 vs GPT-6.

The SWE-bench Pro gap between them is 5.7 points in Claude’s favour. The hallucination gap on Artificial Analysis’s AA-Omniscience benchmark is 50 points, also in Claude’s favour. GPT-5.5 answers with far fewer tokens and runs close to twice as fast.

Both models are at the frontier. Neither is universally better. The choice is a workload decision.

The benchmarks that actually matter for developers

SWE-bench Pro is the benchmark that deserves the most weight for coding work. It evaluates models on their ability to resolve real GitHub issues in real codebases — not contrived problems, not synthetic code generation prompts, but the kind of tasks a developer would actually hand off to an AI agent. The vendor-reported results from April 2026, as compiled by Vellum:

SWE-bench Pro:
Claude Opus 4.7: 64.3%
GPT-5.5: 58.6%
Gemini 3.1 Pro: 54.2%

A 5.7-point gap on this benchmark is meaningful. It means Claude Opus 4.7 resolves roughly one additional real GitHub issue for every 17–18 that GPT-5.5 handles. Across a month of agentic coding sessions, that difference compounds.

SWE-bench Verified (the 500-problem verified subset):
Claude Opus 4.7: 87.6% (up from 80.8% on Opus 4.6)
GPT-5.5: leads on some sub-categories, notably terminal-heavy tasks
GPT-5.5 holds an edge on Terminal-Bench 2.0: 82.7% versus Claude’s 69.4%

One caveat: Anthropic ran Opus 4.7’s Terminal-Bench score with the Terminus-2 harness and thinking disabled, so the two numbers are not a clean head-to-head. Terminal-Bench tests complex command-line workflows requiring planning, iteration, and tool coordination. For developers building terminal-native tools, CLI utilities, or DevOps automation, GPT-5.5’s 13-point edge here is not a footnote.

The cost calculation — and why the 20% output premium is not the whole story

Both models are $5 per million input tokens. The pricing diverges on output: Claude Opus 4.7 at $25 versus GPT-5.5 at $30 — a 20% premium for GPT-5.5.

GPT-5.5 was explicitly built to be more token-efficient. OpenAI reported roughly 40% fewer output tokens per Codex task than GPT-5.4, and Artificial Analysis measured a similar 40% drop. The comparison that matters here is against Opus 4.7: Artificial Analysis found GPT-5.5 at xhigh effort about 30% cheaper than Opus 4.7 at max effort to run its full Intelligence Index, and GPT-5.5 at medium effort cost about $1,200 against about $4,800 for Opus 4.7 at max, at a similar score.

The honest caveat: token efficiency claims from model providers should be verified against your specific workloads. OpenAI’s figures come from their internal evaluations, which may not reflect the distribution of tasks in your codebase. Claude Opus 4.7 is verbose — it explains, documents, and narrates as it works, which is useful in review contexts and expensive in agentic loops. If you are building an agentic system where the model’s output goes directly to execution rather than to a developer for review, verbose output is pure cost with no benefit.

The practical cost estimate for a mid-volume agentic coding use case:

Assume 500 API calls per day, 2,000 input tokens and 1,000 output tokens per call on Claude Opus 4.7. Monthly:

  • Input: 30M tokens × $5/1M = $150
  • Output: 15M tokens × $25/1M = $375
  • Total: $525/month

If GPT-5.5 produces the same work with half the output tokens (500 per call):

  • Input: 30M tokens × $5/1M = $150
  • Output: 7.5M tokens × $30/1M = $225
  • Total: $375/month

That is about 29% cheaper, in line with Artificial Analysis’s measured gap. If GPT-5.5 only matches Opus 4.7’s output length, it costs more, so the efficiency has to hold for your task distribution. The efficiency is real on well-scoped tasks with clear outputs. It is less pronounced on complex architectural reasoning where the model needs to think through tradeoffs before responding.

The reducing Claude API costs post covers prompt caching and context management approaches that apply regardless of which model you choose — those optimisations reduce the input token cost, which is identical for both models.

Where Claude Opus 4.7 is the clear choice

Long-horizon autonomous work. Complex refactors spanning multiple files, agentic sessions that run for extended periods, tasks where the model needs to track state across many tool calls — Claude Opus 4.7 is more reliable in these contexts. Anthropic built the model for exactly this, and its launch post quotes Rakuten reporting that Opus 4.7 resolves 3x more production tasks than Opus 4.6 on its internal Rakuten-SWE-Bench. The new xhigh effort setting (between “high” and “max” in Opus 4.7) gives developers granular control over reasoning depth versus latency.

Code review with a reliability requirement. The hallucination rate gap on AA-Omniscience, 36% for Claude and 86% for GPT-5.5, measures how often each model answers instead of admitting it does not know. The Decoder noted that GPT-5.5 still posted the highest accuracy on that benchmark, so it knows more but guesses more. That guessing is what should give developers pause for production code review. Incorrect information in a code review that gets missed costs more than the $5/million token savings. For any use case where false positives or fabricated findings cause real problems, Claude Opus 4.7’s accuracy is worth the cost difference.

Large codebase comprehension. Claude’s 1 million token context window, combined with its architectural reasoning benchmarks, makes it the stronger choice when you are asking the model to understand a large system before modifying it. The agentic coding guide covers why context coherence matters in these workflows — and it does, especially when the model needs to track how a change in one module affects five others.

Where GPT-5.5 is the clear choice

High-volume API pipelines. If you are running thousands of calls per day on well-defined tasks — code explanation, docstring generation, type annotation, test scaffolding — GPT-5.5’s token efficiency advantage translates directly to operating cost reduction. At scale, Artificial Analysis’s roughly 30% per-task saving is a meaningful budget difference.

Terminal-heavy and DevOps automation. The 82.7% Terminal-Bench 2.0 score versus Claude’s 69.4% is a 13-point gap that shows up in practice: GPT-5.5 is better at planning and executing complex shell workflows, debugging build pipelines, and navigating file systems in terminal contexts. If your agentic use case involves a lot of shell interaction, GPT-5.5’s training shows.

Speed-sensitive interactive tools. Artificial Analysis measures GPT-5.5 (xhigh) at about 99 tokens per second versus about 53 for Claude Opus 4.7 (max). For tools where developer response latency matters — interactive pair programming, real-time code suggestion, anything where you are waiting for the model — GPT-5.5’s speed advantage is perceptible.

OpenAI ecosystem integration. If you are already deeply integrated with OpenAI’s ecosystem — using Codex CLI, Responses API, or building on their tooling — switching to Claude for frontier model access adds API management complexity. GPT-5.5 is the straightforward choice when ecosystem consistency has value.

The pricing precedent worth noting

GPT-5.5 doubled both prices from the previous GPT-5.4 line, from $2.50/$15 to $5/$30 per million tokens. This happened 23 days after GPT-5.4 launched. The precedent for rapid repricing is now established at OpenAI: a new model version means a new pricing tier, and the transition can happen faster than many engineering teams’ planning cycles.

Claude Opus 4.7 launched at the same price as Opus 4.6 ($5 input / $25 output), continuing Anthropic’s practice of pricing model upgrades at equivalent or lower rates than the versions they replace. Whether this pricing discipline continues is worth monitoring — but the current data point favours predictability on the Claude side.

The practical decision

If you are building an autonomous coding agent and reliability is the top priority: Claude Opus 4.7. SWE-bench Pro leadership and a 36% hallucination rate are the numbers that matter most for agents that operate with limited human review.

If you are running high-volume pipelines and operating cost is the constraint: GPT-5.5 — but verify the token efficiency claim on your actual workloads before committing.

If you are doing terminal automation and DevOps tooling: GPT-5.5. The 13-point Terminal-Bench gap is real.

If you need to make a decision without running your own evaluation: Claude Opus 4.7’s benchmark profile is more consistently suited to the tasks that developers actually run in production. GPT-5.5 wins in specific niches that are genuinely important — but those niches require you to know which one you are in.


SWE-bench Pro and Terminal-Bench 2.0 figures are vendor-reported, compiled by Vellum. Hallucination rates and token efficiency from Artificial Analysis. Pricing from OpenAI’s model page and Anthropic’s launch post.

(Updated October 12, 2026: corrected the token-efficiency claim from “72% fewer output tokens than Opus 4.7” to OpenAI’s roughly 40% fewer than GPT-5.4, with Artificial Analysis’s measured ~30% per-task saving versus Opus 4.7; corrected output speed from 74 vs 50 to about 99 vs 53 tokens per second per Artificial Analysis; attributed the 36% vs 86% hallucination rates to AA-Omniscience and clarified what they measure; corrected the “3x more production tasks” figure to a Rakuten customer quote; added that GPT-5.4 input was $2.50; noted that both models have been superseded by Claude Opus 5.5 and GPT-6.)