Claude vs GPT vs Gemini: Coding Benchmark Leaderboard

A June 2026 coding benchmark comparison of Claude, GPT, and Gemini for model buyers.

On the table are Claude, GPT, and Gemini, and this comparison helps you decide which one fits coding work when price, context, and benchmark evidence do not line up cleanly.

At a glance

Claude: strongest on breadth, weaker on public benchmark visibility

Claude Fable 5 and Claude Opus 4.8 are the least tidy options to compare because Anthropic’s public coding tables were not machine-readable on the verification date. That does not mean they are weak models. It means the leaderboard-style proof is harder to extract from the source, so buyers have to lean more on the spec sheet, the product tier, and their own tests.

The practical distinction is price and positioning. Claude Opus 4.8 is the better value inside Anthropic’s lineup at $5 per million input tokens and $25 per million output tokens, with a 1M-token context window and 128K output. Fable 5 doubles that to $10 / $50, which signals a premium tier for teams that want Anthropic’s top release even when the benchmark table is not easy to quote back.

GPT: a strong middle ground if you want long context and a readable score

GPT-5.5 is the cleanest OpenAI option in this comparison because the pricing and context specs are easy to verify, and Anthropic’s page attributes a Terminal-Bench 2.1 score of 83.4% to it. The caveat is important: that number is competitor-reported, not a score we read directly from OpenAI, so it should be treated as directional rather than final proof.

Still, GPT-5.5 looks attractive for teams that want a large 1,050,000-token context window, 128K max output, and a familiar $5 / $30 price point. If your coding workflow involves long repo-wide prompts, agent loops, and lots of pasted context, GPT-5.5 gives you a lot of room without moving into the most expensive tier.

Gemini: best verified public scores and the lowest entry price

Gemini 3.1 Pro is the most benchmark-transparent model in this set because Google’s official model card publishes the figures directly: 80.6% on SWE-bench Verified, 54.2% on SWE-bench Pro Public, and 2887 Elo on LiveCodeBench Pro. Those are single-attempt results from the official card, which makes them especially useful if you care about public, source-backed numbers more than vendor-adjacent references.

It also has the sharpest price advantage at $2 / $12 per million tokens for prompts up to 200K, with higher tiers above that, but the trade-off is a 64K max output and a benchmark profile that is not directly comparable across every test harness. Gemini is the value play when you want a strong published score, lower cost, and you can live with shorter outputs.

When to pick what

Pick Claude Opus 4.8 if you want the safest Anthropic default for everyday agentic coding and you care more about a balanced price-to-capability mix than a headline benchmark citation.

Pick GPT-5.5 if your team works in very long contexts, wants a large output budget, and prefers a model with a strong but caveated benchmark signal that still reads well in vendor comparisons.

Pick Gemini 3.1 Pro if you want the lowest cost, the most clearly published public coding scores, and a model that is easy to justify in a procurement review because the source numbers are directly visible.

Default to Gemini 3.1 Pro for cost-sensitive coding teams, but switch to GPT-5.5 when your workflows regularly need the extra-long 1,050,000-token context window and higher-output headroom.