· 5 min · Tools
Anthropic Ships Claude Opus 5 at Half the Price of Its Own Flagship
Opus 5 costs the same as the model it replaces, half what Claude Fable 5 costs, and outscores Fable 5 on agentic coding — while still trailing it by 0.8 points on SWE-bench Pro.
Claude Opus 5 shipped on July 24 at $5 per million input tokens and $25 per million output tokens — exactly the price of Claude Opus 4.8, the model it replaces, and half of what Anthropic charges for Claude Fable 5, which the company still calls its most capable model.
That combination is the story. Tokens are the chunks of text a model reads and writes — roughly three-quarters of a word each — and they're the unit API access is billed in. The pattern at Anthropic through the first half of 2026 was that a smarter model meant a costlier one, with Fable 5 at the top charging $10 and $50 per million. Opus 5 breaks that: on the tasks most people are actually buying a frontier model for right now — agentic coding, meaning multi-step engineering work a model carries out on its own without a human approving each step — the cheaper model scores higher than the expensive one sitting above it.
The numbers
On Frontier-Bench v0.1, which measures how often a model completes a multi-step engineering task end to end without human help, Opus 5 passes 43.3% of tasks at maximum effort. Fable 5 manages 33.7%. Opus 4.8, the model Opus 5 directly replaces at the same price, manages 18.7%.
The picture is less lopsided on the older coding benchmarks. On SWE-Bench Verified — a curated set of real GitHub issues where a model's patch has to make the project's own test suite pass — Opus 5 reports 96.0%, at the top of a leaderboard where the field is now bunched within a couple of points. On the harder SWE-Bench Pro, which uses problems the models are less likely to have seen, Opus 5 scores 79.2% and Fable 5 still edges it at 80.0%. Anthropic is not claiming Opus 5 is its smartest model; that title stays with Fable 5.
Outside coding, the widest gap is on ARC-AGI-3, a set of puzzle-style tasks built specifically to resist memorization. Opus 5 scores 30.2%, against 7.8% for OpenAI's GPT-5.6 Sol and 1.5% for Opus 4.8 — a jump large enough that it's worth treating as provisional until independent evaluators reproduce it.
| Input / output per 1M tokens | SWE-Bench Pro | Frontier-Bench v0.1 | |
|---|---|---|---|
| Claude Opus 5 | $5 / $25 | 79.2% | 43.3% |
| Claude Fable 5 | $10 / $50 | 80.0% | 33.7% |
| Claude Opus 4.8 | $5 / $25 | 69.2% | 18.7% |
The dial Anthropic is selling
Much of the launch coverage, including Fortune's, frames Opus 5 as introducing an effort setting that trades intelligence for cost. That framing overstates it. The effort parameter has been generally available since the Claude 4.6 family and runs across five levels — low, medium, high, xhigh, and max — on Opus 4.7 and 4.8 as well. What's new is the emphasis: Anthropic is now pitching the dial as the headline answer to enterprise buyers whose token bills became impossible to forecast, and Opus 5 holds up unusually well at the cheap end of it, where earlier models degraded sharply.
There's also a fast mode, which runs the same model roughly 2.5× faster at double the price — $10 and $50 per million, which lands it back at Fable 5's rate.
What actually breaks
Swapping the model string to claude-opus-5 is not a no-op, and two changes will bite quietly rather than loudly.
Thinking is now on by default. On Opus 4.8 and 4.7, omitting the thinking parameter meant the model didn't reason before answering; on Opus 5 the same request runs adaptive thinking. Because max_tokens caps thinking and response text together, a call that was sized tightly around its answer can now truncate mid-sentence with nothing in the error to explain why.
Turning thinking off is also newly constrained: thinking: {"type": "disabled"} returns a 400 error when effort is set to xhigh or max. It's accepted at high or below. The check runs per request, so a route that raises effort partway through a conversation fails even though earlier calls succeeded.
Two smaller changes are worth knowing. The minimum cacheable prompt drops to 512 tokens from 1,024 — prompt caching bills repeated context at roughly a tenth of the normal rate, so prompts previously too short to qualify now cache with no code change. And Opus 5 draws on a separate rate-limit pool from the combined Opus 4.x bucket, so moving traffic over neither frees headroom on the old limit nor inherits it.
Who gets it
Opus 5 is available through the Claude API, claude.ai, Claude Code, and Claude Cowork. It's the default model on Claude Max and the strongest model available on Claude Pro — which lands four days after Anthropic folded Fable 5 back into Max and Team Premium at half the plan's usage cap, and less than three weeks after Pro subscribers were moved to metered usage credits for Fable 5 access. For a Pro subscriber, the practical effect is that the model they can reach without paying per token just got considerably better.
Anthropic also says Opus 5's cybersecurity classifiers — the filters that decline requests touching offensive security work — intervene around 85% less often than Fable 5's, a direct response to the false positives that made Fable 5 awkward for legitimate security tooling.
What to do with this
If you're on Opus 4.8, this is a straight upgrade at identical pricing, and the only two things that will break are the thinking defaults above. Audit every call site that never set thinking or that pairs disabled thinking with high effort, then swap the string.
If you're paying for Fable 5, run your own evaluation before assuming you still need it. Fable 5 keeps a narrow lead on SWE-Bench Pro and remains the stronger model for the longest autonomous runs, but on agentic coding at half the price the comparison now favors Opus 5 clearly enough that the burden of proof has flipped.
And run an effort sweep rather than inheriting a setting. Start at xhigh for coding and agentic work, high elsewhere — then step down. Opus 5's low and medium are strong enough that the effort level tuned for a previous model is probably costing more than it's buying.