GPT-5.5 vs Claude Opus 5: The Coding Race Nobody Can Score Fairly

GPT-5.5 and Claude Opus 5 are the two models most teams actually mean when they say "frontier coder," and on paper they are close enough that the decision usually gets made on a benchmark table. That is a mistake this year for a reason that has nothing to do with either model's ability: GPT-5.5 is being retired from ChatGPT, ChatGPT Work and Codex on October 14, 2026, and Codex is the surface a large share of its coding users actually work in. Claude Opus 5 is not going anywhere. So the first question in this matchup is not which model writes better code — it is which one still answers the request you are about to send in November.

That reframes the comparison, but it does not settle it. If you call GPT-5.5 through the API or through a Codex session authenticated with your own API key, the retirement does not touch you, and the capability question is exactly as live as it was in April. What follows is the honest version of that question, including the part most comparison pages skip: both models' headline numbers are measured in ways that flatter whoever is running the test.

The one-line answer

If you need the shortest defensible version: Claude Opus 5 is the safer default for production coding work in late 2026, because it is cheaper on output, independently benchmarked, and not being retired from the tool you probably use it in. GPT-5.5 remains the stronger choice on specific fronts — terminal-native agentic work, long-context retrieval, and token efficiency on hard tasks — and on the API path it keeps all of those. Everything below is the reasoning behind that split, and the caveats that make it less clean than it sounds.

Two flagships that launched into different worlds

GPT-5.5 arrived on April 23, 2026 as OpenAI's frontier model for complex professional workloads, built on GPT-5.4 with stronger reasoning, better reliability and — the claim OpenAI led with — improved token efficiency on hard tasks. It accepts text, image and file input, serves roughly 922K input and 128K output tokens, and prices at $5.00 per million input and $30.00 per million output, with cached reads at $0.50 and a higher rate above 272,000 input tokens.

Claude Opus 5 is Anthropic's flagship for demanding reasoning, coding and long-horizon agentic work, and it is tuned for a specific weakness that has dogged agentic coding: staying coherent across very long, multi-step sessions with heavy tool use. It serves a 1M-token context with up to 128K output, accepts text, image and file input, and prices at $5.00 input and $25.00 output with cached reads also at $0.50. It replaces sampling parameters like temperature and top_p with a reasoning-effort dial, which is a real difference in how you tune it but not in what it can do.

The spec contrast

• Price — GPT-5.5 $5.00/$30.00 per 1M tokens vs Claude Opus 5 $5.00/$25.00 per 1M; identical on input, Claude Opus 5 about 17% cheaper on output

• Cached input — both $0.50 per 1M tokens, a 90% discount on the input rate

• Context — GPT-5.5 ~922K input / 128K output vs Claude Opus 5 1M input / 128K output

• Input modalities — GPT-5.5 text, image, file vs Claude Opus 5 text, image, file; the two are at parity here

• Tuning — GPT-5.5 reasoning-effort settings from none to xhigh vs Claude Opus 5 adaptive reasoning via an effort dial, no temperature or top_p

• Availability — GPT-5.5 leaves ChatGPT, ChatGPT Work and Codex on October 14, 2026, but not the API vs Claude Opus 5 on all Anthropic surfaces with no announced retirement

Where the benchmark claims stop being comparable

Start with the independent picture, because it is the only part of this that anyone can check. Artificial Analysis scores Claude Opus 5 at 51 on its Intelligence Index. It scores GPT-5.5 at 39 on the current v4.3 index — and, more tellingly, it has flagged GPT-5.5 as deprecated and wound its evaluation back to default 10,000-token workloads. That is worth reading carefully. The independent record for GPT-5.5 is not just weaker; it is closing, and once it stops, every number still circulating for that model is a vendor number.

On the vendor-reported side, OpenAI's own figures put GPT-5.5 at 82.7% on Terminal-Bench 2.0, 82.6% on SWE-bench Verified and 58.6% on SWE-bench Pro. Anthropic's figures for Claude Opus 5 include 90.2 on Terminal-Bench 2.1, 74.7 on SWE-bench Pro and 68.0 on JobBench. The tempting move is to stack those side by side and declare a winner. Do not — and notice that most comparison pages do it anyway.

The Terminal-Bench line is the clearest example. GPT-5.5's 82.7% is on Terminal-Bench 2.0; Claude Opus 5's 90.2 is on Terminal-Bench 2.1. Those are different benchmark versions with different task sets, and a cross-version comparison is not a comparison. The SWE-bench Pro rows are at least the same suite, and there Claude Opus 5's 74.7 against GPT-5.5's 58.6 is a genuine and wide gap — but both figures are vendor-reported on both sides, and neither has been reproduced under identical harness conditions.

The harness problem, which is the real story

If there is one finding worth carrying out of this matchup, it is that harness choice moves these models more than the gap between them does. An arXiv evaluation found GPT-5.5 performed strongest under Claude Code and then dropped 21.8 points when moved to OpenHands, while Claude Opus 4.8 — the comparison point in that study — fell only 5.5 points on the same change. A 21-point swing from scaffolding alone is larger than most published head-to-head differences between the two flagships.

There is a practical version of this that OpenAI's own framing concedes: GPT-5.5 reaches 83.4% on Terminal-Bench using its native Codex CLI harness, against 72.1% on a standard harness. If your evaluation runs GPT-5.5 outside Codex, you are measuring a different model than the one in the launch table.

This matters for the retirement in a way that is easy to miss. If your GPT-5.5 numbers were produced inside Codex on a ChatGPT sign-in, and that path closes on October 14, you are not just losing a model — you are losing the harness your benchmark is calibrated to. Re-baseline on the API path before you conclude anything about what the successor costs you.

Cost per task, not cost per token

The per-token comparison is straightforward and slightly favours Claude Opus 5: same $5.00 input, $25.00 versus $30.00 on output. The interesting comparison is cost per completed task, and there the picture inverts, because OpenAI's central GPT-5.5 claim was token efficiency — roughly 40% fewer tokens on equivalent tasks by several developer accounts. If a model finishes the same job with 40% fewer output tokens, a $30 output rate can produce a smaller bill than a $25 rate on a more verbose model. That claim is vendor-framed and has not been independently reproduced, so treat it as a hypothesis to test on your own workload rather than a settled fact.

Both models are in our catalogue, which makes that test cheap to run properly rather than anecdotally. Both sit behind one key at provider list rate with 0% markup, so a straight A/B on your own traffic is a model string in the request body, not a second contract — and any vendor repricing lands on our side the same day it is announced. The routing DSL is the useful part here: split live traffic between the two, hold the split at whatever ratio your error budget tolerates, and let the cost-per-task numbers come from production instead of a benchmark page.

What this means for the next six weeks

For a team choosing today, the decision has an asymmetry that a pure capability read would hide. Claude Opus 5 is the lower-risk pick: independently scored, cheaper on output, and not scheduled for removal from the products it is sold in. GPT-5.5 is not a bad model — it is a strong one with a genuine edge in terminal-native agentic work — but choosing it on the subscription path means choosing a model with a known expiry date six weeks out, and on the API path means accepting a thinner and shrinking independent record.

The defensible position is to keep GPT-5.5 where it demonstrably wins and where you control the path — the API, an API-key Codex session — and to make Claude Opus 5 the default everywhere else. What should change your mind in either direction is neither model's marketing: it is the first neutral head-to-head run under a single fixed harness, and whatever OpenAI publishes about whether the API path for GPT-5.5 survives past the product retirement.

Both sides of this matchup sit behind one key: GPT-5.5 and Claude Opus 5 are a model string apart, with automatic failover between them.