Gemini 3.1 Pro vs Claude Opus 4.7 vs GPT-5.5: Benchmarks, Pricing, and Which to Pick

I pulled the verified numbers from Anthropic's Opus 4.7 announcement, OpenAI's GPT-5.5 page, Google's Gemini 3.1 Pro post, Vellum's deep dives on each model, Artificial Analysis's cross-vendor Intelligence Index, and the Hacker News threads where developers stress-tested the launches. Then I cross-checked pricing on OpenRouter. Where vendors disagree, I'll say so.

• Shipping production code → Claude Opus 4.7. 87.6% SWE-bench Verified, 64.3% SWE-bench Pro, 77.3% MCP-Atlas. Cursor measured a 12-point IDE-workflow jump from Opus 4.6 to 4.7 on its internal CursorBench (58% → 70%).
• Long-context research over >200K tokens → Gemini 3.1 Pro. 1,048,576-token window, 85.9% on BrowseComp, and Opus 4.7 just regressed on long-context retrieval to 59.2% from 91.9%.
• Agentic terminals, math, ARC-AGI, cost-per-intelligence → GPT-5.5. 82.7% Terminal-Bench 2.0 (+13.3 over Opus), 85.0% ARC-AGI-2, 60 on AA Intelligence Index, ~$1,200 to run the full suite.
• Anything where confidence-without-knowing is unacceptable (legal, medical, compliance) → Opus 4.7. Its 36% AA-Omniscience hallucination rate is less than half of GPT-5.5's 86%.
• Daily writing → Honestly? None of these three. Opus 4.7 has a documented writing-quality regression (more on that below). Gemini 3.1 Pro is awkward in tool-driven UIs. GPT-5.5 hallucinates. If writing is the job, GPT-5.4 or Claude Sonnet 4.5 may still be the right call.

Three takeaways jump off this table. First, GPT-5.5 owns terminals, agents-in-shells, abstract reasoning, and frontier math — the workloads where step-by-step planning matters more than calibrated knowledge. Second, Opus 4.7 owns the IDE coding stack. Third, the GPQA Diamond scores (93.6% / 94.2% / 94.3%) are statistically indistinguishable — graduate-level Q&A is no longer a useful differentiator at the frontier.

At 1M input + 1M output tokens per day under 200K-token prompts, the napkin math is brutal: Opus 4.7 = $30/day = $900/month. GPT-5.5 standard = $35/day = $1,050/month. Gemini 3.1 Pro = $14/day = $420/month. Over 30 days, going Gemini saves you $480 vs Opus. Going Opus over GPT-5.5 saves you $150 — and gets you the lower hallucination rate as a bonus.

Then there's Artificial Analysis's like-for-like Intelligence Index run: ~$1,200 on GPT-5.5 medium, ~$4,800 on Opus 4.7 max, ~$900 on Gemini 3.1 Pro Preview. Opus 4.7 max costs 5.3× as much as Gemini for a statistically equivalent Intelligence Index score (57 vs 57). That's the cost-per-intelligence story Anthropic doesn't put on its pricing page.

The official case for Opus 4.7 on coding is strong. SWE-bench Verified at 87.6%, SWE-bench Pro at 64.3% (+5.7 over GPT-5.5), MCP-Atlas at 77.3%, and Finance Agent v1.1 at 64.4%. Cursor's internal CursorBench jumped from 58% on Opus 4.6 to 70% on Opus 4.7. Notion reported a 14% improvement on its evals and called Opus 4.7 "the first model to pass implicit-need tests." Anthropic's own Rakuten-SWE-Bench number claims 3× more production tasks resolved versus 4.6.

And then you read the Hacker News thread. User agentseal ran three days of side-by-side production coding and posted: "4.7 gets things right on the first try less often than 4.6. One-shot rate sits around 74.5% vs 83.8%. Roughly double the retries per edit (0.46 vs 0.22). Produces a lot more output per call, about 800 tokens vs 372 on 4.6. Cost per call is $0.185 vs $0.112." Multiple users confirm the verbosity regression — one notes 4.7 "reach[es] ChatGPT levels of verbosity in code and loves to overcomplicate the most simple things."

Anthropic shipped a new tokenizer with 4.7 that makes the same input map to 1.0–1.35× more tokens. The list price didn't move; the bill can. Combine that with longer outputs at the new default xhigh effort level — the recommended mode for Claude Code — and you can quietly burn 2–3× more of a Claude Pro 5-hour quota per task. User hgoel: "Each exchange takes ~5% of the 5-hour limit now, when it used to be maybe ~1-2%."

This branch flipped at the model-card level. Anthropic's own Opus 4.7 model card lists 59.2% on long-context retrieval, down from 91.9% on Opus 4.6 — a 32.7-point regression. Hacker News user bachittle surfaced the number directly from the card. User Someone1234 confirmed: "Opus 4.7 is also worse at 256K context. Across the board regression, not just 1M context." If your workflow lives in the 200K–1M token band — repository-scale code review, contract corpora, multi-PDF research synthesis — Opus 4.7 is now the wrong tool.

Gemini 3.1 Pro inherits the slot by default. The window is 1,048,576 tokens with a 65,536-token output cap. BrowseComp leads at 85.9%. MMMLU multilingual at 92.6% beats both rivals — useful if your corpus isn't English. Artificial Analysis benchmarks output speed at 115.5 tokens/sec.

The honest negatives matter. Time-to-first-token sits at a slow 20.87 seconds — bad for interactive chat, fine for batch research. Pricing doubles above 200K tokens to $4 input / $18 output per million, which makes a single 800K-token research run materially more expensive than it looks at the headline rate. And independent evaluator Dan Cleary called Gemini 3.1 Pro "the smartest dumb model I know" after watching it choke on a ChatGPT-clone build that Claude Sonnet 4.6 "handled effortlessly" — meaning the long-context win does NOT translate to UI-driven coding.

GPT-5.5 has the highest score on the Artificial Analysis Intelligence Index (60), Terminal-Bench 2.0 (82.7%, +13.3 over Opus), OSWorld-Verified (78.7%), ARC-AGI-2 (85.0%), FrontierMath Tier 4 (35.4%, +12.5 over Opus), MRCR v2 across 512K–1M (74.0% vs Opus's 32.2%), CyberGym (81.8%), and GDPval knowledge-work tasks (84.9%). It's also natively omnimodal — text, image, audio, video — where the other two require separate handoffs for some media types. At ~$1,200 to run the full Intelligence Index suite, it's mid-priced between Gemini and Opus while delivering the highest measured intelligence score.

If you're building agentic pipelines that drive a shell, hit OS-level UI elements, or chain 50+ tool calls, GPT-5.5 is the right default. The dossier's vendor-partner data backs this up: OpenAI's claim that GPT-5.5 "matches GPT-5.4's per-token latency in real-world serving" while using ~40% fewer output tokens turns the doubled output price into roughly a 20% effective increase on Codex-style work.

Now the asterisk. GPT-5.5's AA-Omniscience hallucination rate is 86%. Opus 4.7 is 36%. Gemini 3.1 Pro is 50%. Jake Handy of Handy AI didn't mince words: "the hallucination gap is insane — at 86%, the model will confidently answer more questions it doesn't know the answer to, making it unsuitable for medical, legal, or regulatory work where accuracy is essential." Stephen Smith framed the same finding for legal practitioners: "GPT-5.5 may know more, reason better, and still be more willing to make something up when it should say 'I don't know.' For legal research, that is not a footnote. That is the caveat."

OpenAI argues hallucinations are 60% lower than GPT-5.4 (per Mindwired AI's roundup of OpenAI's own claims). Third-party verification is still in progress, and the AA-Omniscience absolute number is what it is.

1. Opus 4.7's tokenizer change. Same prompt, 1.0–1.35× more tokens. Sticker price unchanged means a stealth 0–35% bill increase depending on input mix. Combined with longer default outputs at the new xhigh effort level, real-world cost-per-task is up.
2. Gemini 3.1 Pro's 200K cliff. $2/$12 per million is the marketing rate. Cross 200,000 input tokens and you're at $4/$18 — a 100% jump. A single repository-scale research prompt can quietly land in the doubled tier.
3. Opus 4.7's adaptive thinking can't be disabled via the standard API parameter. Per Hacker News user rkuska, the manual thinking parameter is rejected on claude-opus-4-7 with a 400 error. User JamesSwift: "Disabling adaptive thinking plus increasing effort seem to be what has gotten me back to baseline performance." If you scripted reasoning controls before, they need a rewrite.
4. Opus 4.7 ships with sharper instruction-following and a writing regression. Per botmonster's reception summary, "Opus 4.7 follows instructions much more literally than 4.6, which means prompts that worked before can now produce unexpected results." Hacker News user limalabs: "Man is it bad at writing. It's such a stark contrast, sloppy, unprecise, very empty sentences." One reviewer at BoringBot scored Opus 4.7 at 35/50 on PRD writing where 4.6 hit 45/50, with the model "literally stopping mid-sentence" against token caps.
5. GPT-5.5's effective cost depends on your workload mix. The "40% fewer output tokens" efficiency claim is real on Codex-style coding. It's not validated for long-form writing or research. Run your own A/B before assuming the headline efficiency.

The honest summary: if you can only license one of these three for your team this quarter, license Claude Opus 4.7 — it has the lowest hallucination rate, leads the coding benchmarks that map to actual product work, and its weaknesses (long-context retrieval, writing, verbosity) are well-mapped enough that you can route around them. If you do enterprise-scale long-document analysis, add Gemini 3.1 Pro as the second seat — the <200K pricing makes it cheaper than the alternatives and the long-context numbers aren't close. Add GPT-5.5 only if you're building autonomous agents that live in shells, or doing math-heavy research where calibration matters less than reasoning ceiling.

For the broader landscape — including Chinese open-source models that beat all three on price — see our task-by-task model picker guide and the prior-generation comparison in GPT-5.4 vs Opus 4.7 vs Gemini 3.1 Pro. The frontier moves fast; the decision framework above is what's true on April 27, 2026.