Claude Opus 5 review: a flagship built for agentic coding

Claude Opus 5 (anthropic-claude-opus-5) is the model to reach for when a hard problem has to be solved and quality matters more than cost. It ranks near the top in math, hard prompts and instruction following. Its weak spot is just as clear: in long multi-turn conversations it ranks much lower.

What it is

Opus 5 is Anthropic's newest Opus and the most capable model in the Opus line.

• Context window: 1M tokens (the default and the maximum)
• Max output: 128K tokens
• Input: text, images, files
• Thinking: on by default, scaled to the task without any setting

Anthropic's Claude Fable 5 sits a class above it. Opus 5's appeal is flagship-level intelligence one price tier down.

Reputation: long agentic runs and code review

What Opus 5 gets the most credit for is agentic development work that runs on its own for a long time. It is tuned to finish the job instead of stopping at stubs and placeholder code, and Anthropic positions it as more consistent than Claude Opus 4.8 over long tasks.

That is the vendor's positioning and users' reports. The leaderboard data below says something different about multi-turn work, so check it against your own workflow.

Code review and bug hunting are the other strength. Most of the bugs it reports are real, and it is credited with good precision (few false positives) and recall (few misses) at the same time. A practical setup is a cheap, fast model for the first pass and Opus 5 for a careful second one.

Strengths and weak spots, by category

Third-party benchmarks were thin at launch, but arena-style leaderboard data has since come in. These are the category ranks FastMetal had imported as of July 31, 2026, for claude-opus-5-high on the text board.

The leaderboard lists effort variants (-high, -max) as separate entries, and neither maps one-to-one onto the anthropic-claude-opus-5 FastMetal serves, which runs at the default effort. Read the table for the pattern, not as an exact score.

The same model is 2nd in math and 19th in multi-turn. The overall 7th averages that spread away.

The takeaway: very strong on a single hard problem, relatively weaker when a conversation builds up over many turns. If you plan to hand it long autonomous sessions, test that on your real workflow first.

The gaps are also smaller than the ranks suggest. Between 1st and 20th on this table, the Elo difference works out to roughly a 60 percent win rate; the top of the board is a pack.

On the newer September 25, 2026 board, the same -high entry was 11th overall of 409 models and 9th on hard prompts of 409.

high vs max

The effort variants are good at different things. Same July 31 data:

high for math and code, max for creative writing. More effort is not better everywhere.

Cost

Opus is a premium tier, billed per million input and output tokens. Opus 5 costs the same as Opus 4.8 and less than Fable 5; current prices are on the Claude Opus 5, Claude Opus 4.8 and Claude Fable 5 model pages. Long agentic runs consume a lot of tokens in total, so estimate cost per task rather than reading the per-token rate alone.

FastMetal is prepaid: you set the ceiling before you start, so a runaway agent cannot run up an open-ended bill.

Japanese

If your output is in Japanese, the ranking moved. On the July 31 board Opus 5 was not in the top ten of the Japanese category, which was led by Claude Fable 5, the GPT-5.5 family and the Gemini 3 Pro family. On the September 25 board, claude-opus-5-high was 8th of 271 in Japanese. The best models for Japanese page has the current order.

Good fit, poor fit

Good fit

• Large features and refactors that span many files
• Long, multi-step autonomous agent runs (after you have checked the multi-turn behavior)
• Code review, bug hunting, debugging
• Reading a whole repository into a 1M-token context

Poor fit

• One-line fixes, simple classification and other jobs a small model handles; the cost doesn't pay off
• Anything where the only goal is high volume at low cost

FAQ

Is it worth moving from Opus 4.8?

If your work is mostly long autonomous coding and bug hunting, the gains in consistency and precision should show. The price is the same, so run the same prompts on both and compare.

How is it different from Fable 5?

Fable 5 is the class above, at a higher price. Opus 5 is flagship-level intelligence for less.

Should I use high or max?

It depends on the task. On the July 31 data, high ranked higher in math and coding and max in creative writing.

Is it as good at long agentic sessions as its reputation says?

It is very strong on single hard problems, but it ranked 19th in the multi-turn category. Test it on your own workflow before you hand it long autonomous runs.

Try it

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic-claude-opus-5",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

On FastMetal, anthropic-claude-opus-5 and the smaller models share one API key and one balance. Change model and you can compare quality, cost and speed with the same code. Head-to-heads: Claude Opus 5 vs GPT-5.6 Sol, DeepSeek V4 Pro vs Claude Opus 5 and Qwen3.8 Max vs Claude Opus 5.

Whether to adopt it is best decided by an eval on your own tasks, not by a public leaderboard.