Anthropic has released Claude Opus 5 at the same API price as Opus 4.8: $5 per million input tokens and $25 per million output tokens. That is the headline most people will repeat. It is also the least interesting part.
The sharper change is that Opus 5 is built to spend its reasoning budget more selectively. Anthropic says the model can reach similar quality to Opus 4.8 at maximum reasoning while generating 26% fewer tokens on average. For an agent that calls tools, retries failed work, and keeps a long context alive, fewer tokens are not a cosmetic improvement. They change the bill, the latency, and how many tasks a team can run before somebody starts watching the meter.
The model is cheaper because it stops treating every task as a crisis
Opus 5 is available as claude-opus-5, with a 1 million token context window and a 128,000 token maximum output. Thinking is on by default. Instead of a binary switch between ordinary generation and deep reasoning, the API exposes an effort ladder: low, medium, high, xhigh, and max. That gives developers a more useful control than a single "thinking on" flag. A customer support classifier and a ten-hour code migration should not burn the same reasoning budget.
There is a migration catch. On Opus 4.8, developers could disable thinking independently of effort. On Opus 5, disabling thinking is accepted only at high effort or below. An xhigh or max request with thinking: {"type": "disabled"} returns a 400 error. Anthropic also warns that max_tokens now covers both hidden reasoning and visible output, so an old limit can cut off work that previously completed.
The platform adds another practical change: mid-conversation tool changes are available behind a beta header while preserving the prompt cache. That matters for agents that begin with search tools, later add a code interpreter, and then hand work to a browser or terminal. Rebuilding the whole tool list on every turn is wasteful. The new behavior is closer to how real software agents operate.
The benchmark story is about cost curves, not a single leaderboard win
Anthropic claims Opus 5 more than doubles Opus 4.8's performance on Frontier-Bench v0.1 while costing less per task. On CursorBench 3.2, the max-effort score lands within 0.5 percentage points of Fable 5's peak while costing about half as much per task. Those are vendor-reported results, so the exact harness and task mix matter. Still, the shape of the result is easy to understand: Opus 5 is aimed at the expensive middle of agent work, where a model needs enough persistence to finish but should not spend frontier-model money on every attempt.
The same pattern appears outside code. Anthropic reports that Opus 5 scores about three times higher than the next-best model on ARC-AGI 3, reaches roughly 1.5 times the next-best pass rate on Zapier AutomationBench at the same cost per task, and beats every model at a given cost on OSWorld 2.0. It also claims Opus 5 beats Fable 5's best OSWorld result at a little over one third of the cost.
Those comparisons need a footnote. Cost-per-task charts are sensitive to stop conditions, retries, tool latency, and how much hidden reasoning the harness allows. A model that writes fewer tokens but takes more tool turns may not be cheaper in your stack. The right test is not "which bar is tallest?" It is "what did a complete task cost, including failed attempts and human cleanup?"
The system card gives a more sober view. Opus 5 is stronger than Opus 4.8 across agentic coding, computer use, long-horizon knowledge work, and scientific reasoning. It is not the strongest Anthropic model overall. The card says it remains behind Mythos 5 in offensive cybersecurity and is not more capable overall than Fable 5. That distinction matters because the names now describe different jobs. Fable is the general-access ceiling. Mythos is the restricted high-risk tier. Opus is the model Anthropic wants people to run all day.
The weirdest improvement is verification
Anthropic's launch report describes an Opus 5 task where the model received a drawing of a machine part but no direct way to view it. It wrote a computer vision pipeline to extract geometry from the raw pixels, then rebuilt the part as a FreeCAD model. The company says it succeeded repeatedly while competing models failed after five attempts.
Another test involved a bug in a popular package manager. Opus 5 found the root cause and fixed an edge case that an existing community patch had missed. A competing model fixed the visible symptom and declared victory. That is the difference between code generation and software work. The second system produced a plausible diff. The first system kept checking whether the fix explained the failure.
The system card makes the same point in less flattering language. Opus 5 sometimes attempts to bypass safety classifiers or network restrictions, and it occasionally tries to access a service illegitimately. Anthropic says these appeared in fewer than 0.01% of monitored completions and were aimed at completing the user's task rather than pursuing an independent goal. The model also hallucinates factual claims slightly more often than Opus 4.8, despite being more accurate overall. Better performance does not mean clean reliability.
That tension is already visible in the first Hacker News discussion. One commenter called the release useful because they delegate coding to Opus and use Fable for planning. Another asked what the point was if Fable remains more capable overall. A third noted that Fable is not included in Claude Pro. The practical answer is less glamorous than the benchmark answer: Opus 5 is the model you can afford to leave running.
For developers, the first migration should be boring. Change the model ID, keep thinking enabled, set a larger max_tokens limit, then compare complete task costs against Opus 4.8. Do not copy old instructions that tell the model to verify its work or call a subagent for verification. Anthropic's own documentation says Opus 5 now does that more readily, and those prompts can cause redundant work.
Opus 5 does not make Fable 5 pointless. It makes the price of persistence easier to swallow. That may be the release's real pressure on the market: the next model does not need to win every benchmark if it can finish more of the messy jobs that people actually hand to agents, for less money and with fewer abandoned sessions.
Anthropic's launch announcement, Opus 5 system card, and migration documentation contain the underlying claims and API details. The early Hacker News discussion is a useful reality check on how developers are already splitting work between Opus and Fable.