benchr
Start with the number that surprises people: these two flagships charge the exact same $5 per million input tokens. That's the price of feeding them your code, your docs, your context. The split shows up on output: Opus 4.8 bills $25 per million, GPT-5.5 bills $30. A dollar-per-million edge sounds tiny until you're generating long agent transcripts all day, where output dominates the bill.
This is the rematch of an older fight. Last time it was GPT-5 against Opus 4.7, and both labs have shipped a point release since. The shape of the contest has changed, so the verdict deserves a fresh look rather than a swapped-in version number.
| Model | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|
| Claude Opus 4.8 | $5 | $25 | 1M context, 128K max output |
| Opus 4.8 fast mode | $10 | $50 | Up to 2.5× output speed, research preview |
| GPT-5.5 | $5 | $30 | 1.05M context, $0.50 cached input |
| GPT-5.5 Pro | $30 | $180 | Higher-accuracy tier, no cached discount |
Coding, where Opus 4.8 pulls ahead
If your decision rests on writing and reviewing software, Opus 4.8 is the pick. Anthropic reports it at 69.2% on SWE-bench Pro, the tougher real-world variant that scores end-to-end GitHub issue resolution, against the 58.6% OpenAI reports for GPT-5.5. That's a wide gap on the benchmark closest to actual production work. Anthropic also calls Opus 4.8 the strongest computer-use and browser-agent model it has tested, at 84% on Online-Mind2Web.
The quieter win is about trust. Anthropic says Opus 4.8 is around four times less likely than Opus 4.7 to let a flaw in code it wrote pass without comment, and more willing to flag when it isn't sure. For code review, that instinct to raise a hand is worth more than a benchmark point, because the bug that ships quietly is the one that costs you. The full picture is in the Opus 4.8 review.
Winner on production coding: Opus 4.8, clearly.
The benchmark Opus loses
Now the honest part, because a comparison that only lists one model's wins isn't worth reading. GPT-5.5 beats Opus 4.8 on Terminal-Bench, the benchmark for long, multi-step command-line agent work. On Anthropic's own run with the common Terminus-2 public harness, Opus 4.8 lands around 74.6% to GPT-5.5's 78.2%. OpenAI reports GPT-5.5 even higher, a state-of-the-art 82.7% on Terminal-Bench 2.0, and Anthropic's footnote notes GPT-5.5 reaches 83.4% under OpenAI's own Codex CLI harness.
The harness matters, and the scores move with it, but the direction doesn't change: GPT-5.5 is the stronger terminal agent. If your workload is an autonomous agent grinding through shell commands, builds, and tool calls for hours, that's GPT-5.5's home turf, and it's exactly the kind of work OpenAI built this release around.
Winner on terminal-agent work: GPT-5.5.
GPT-5.5 as the daily driver
Outside of raw coding, GPT-5.5 is the broader generalist. It carries a slightly larger 1.05-million-token context window, OpenAI tuned it for concise answers, and it's pitched squarely at professional, document-heavy knowledge work: research, synthesis, analysis. It's also the model behind ChatGPT's default for everyone, so it's the one most of your non-developer colleagues are already using. The GPT-5 review traces how that generalist lineage holds up.
Where this really bites is writing. Neither lab markets a "best writer," and for long-form drafting the two trade blows in ways benchmarks won't capture. We pull that specific contest apart in Claude vs ChatGPT for long-form writing, including the output-length ceilings that decide how much either one can produce in a single pass.
Fast mode and the cost math
Opus 4.8's fast mode is the new lever. For $10 input and $50 output per million, double the standard rate, you get up to 2.5 times the output tokens per second, with identical model behavior. The headline is that this is roughly three times cheaper than fast mode on Opus 4.7, which ran $30 and $150. If latency is your bottleneck on an interactive coding agent, that's a real option now rather than a luxury. It's still gated behind a waitlist and API-only, so treat it as a tool you grow into.
For everything cost-sensitive, the matched $5 input price is the headline you should plan around. Feeding context is free-of-difference between these two; only your output mix moves the bill. For a fuller cost breakdown across workloads, price per use case does the math.
Make Opus 4.8 your default for writing and reviewing code, where it wins SWE-bench Pro, computer use, and the willingness to flag its own mistakes. Switch to GPT-5.5 for terminal-agent runs, the broadest knowledge work, and as the all-purpose model your whole team can share. Their input prices match, so go with both if you ship software, and skip the GPT-5.5 Pro tier unless you've measured that you need it.