Anthropic shipped a new flagship today, and the pitch is not what you would expect.

Claude Opus 5 is out, live right now across every Anthropic platform and available as claude-opus-5 on the API. It is the new default on Claude Max and the strongest model Claude Pro subscribers can reach. Pricing did not move: $5 per million input tokens and $25 per million output, exactly what Opus 4.8 cost.

Here is the part worth sitting with. Anthropic is not claiming this is its smartest model. Fable 5 still holds that title. What Anthropic is claiming is that Opus 5 gets close to Fable 5 at half the price, and that this matters more. This is the fourth model the company has shipped in under two months, after Mythos 5, Fable 5, and Sonnet 5, and the whole release reads as a bet that the AI race has stopped being about the frontier and started being about what you can afford to run all day.

Start with the benchmark that Anthropic is leading with. On Frontier-Bench v0.1, an agentic terminal coding evaluation, Opus 5 scores 43.3 percent. Opus 4.8 scored 18.7 percent. That is more than double, and it is well ahead of Fable 5's 33.7 percent, at a lower cost per task.

The rest of the card is consistent. On ARC-AGI 3, which tests novel problem solving the model has not seen before, Anthropic reports Opus 5 scoring three times as high as the next best model. On OSWorld 2.0, a computer-use benchmark, it beats Fable 5's best result at just over a third of the cost. On CursorBench 3.2 at max effort it lands within half a percent of Fable 5's peak score, again at half the cost per task. The one place it does not lead is cybersecurity, where Mythos 5 remains ahead, which is deliberate.

The new control is the effort setting. You can dial the model between low, medium, and high, trading intelligence against speed and token spend. Anthropic's own charts are built around performance at a given cost rather than peak performance, which tells you exactly how the company wants this model judged.

Forget the benchmark tables for a second, because the most convincing evidence came from outside Anthropic entirely.

Opus 5 leaked early, surfacing in Cursor and on Google's Vertex catalog, and developers spent the past few days hammering it before anyone at Anthropic said a word. The demo that spread fastest came from @xikhar, who had it build a Minecraft replica. Recreating Minecraft has become a standard party trick for frontier models, but the details are what made this one land: blocks shattered with correct physics, and light cast real shadows across surfaces. He followed it with an SVG of a PS5 controller where the reflections on the joysticks and touchpad looked closer to a product render than vector art.

The head-to-head is the part worth your attention. Developer @chetaslua gave the identical prompt, a slingshot launching projectiles at a city wall, to both Opus 5 and Fable 5. Fable 5 produced a complete scene in one pass, but its tents rendered as flat blocks of color, the grass had no undulation, and it dropped the parameter panels for projectile tension and arrow weight. Opus 5 kept all of it. He said he had never seen that level of detail out of an Opus model, and enough testers agreed that some started only half-jokingly calling it Fable 6.

Every number above is Anthropic's own. There is no independent verification yet, and the methodology footnotes matter more than usual here. In the Frontier-Bench runs, Opus 4.8 served as the fallback whenever a safety classifier refused a request from Opus 5 or Fable 5, which means the headline comparison includes some assistance from a different model. That is disclosed rather than hidden, which is to Anthropic's credit, but it is the kind of detail that changes how you read a benchmark table.

There is a real feature buried in that same mechanism. When Opus 5 declines a request on safety grounds, the API quietly falls back to another model so you get an answer rather than an error. Useful in production, and also a little strange: a classifier, not you, decides which model answers your prompt. Anthropic says Opus 5's guardrails match Opus 4.8's in most areas, with stronger controls specifically on cyber, and that its cyber ability improved anyway as a side effect of getting generally more capable.

Two practical notes if you are evaluating it today. Opus 5 carries no data retention requirements for general access, which matters if you are under a hard zero-retention policy, and that is a real difference from Fable 5, where Anthropic retains inputs and outputs for safety work. And the efficiency claim is the thing to test yourself, because token-efficiency numbers have a habit of looking different once they meet a real production workload.

Look at what shipped this month and the pattern is hard to miss. OpenAI led its GPT-5.6 launch with token efficiency. Anthropic just launched a flagship whose headline feature is costing the same as the last one while doing twice the work. China's labs are shipping near-frontier open weights at a tenth of Western API prices. Nobody is winning arguments with raw capability anymore, they are winning them with cost per completed task.

For anyone actually building, that shift is good news. The question stopped being whether you can afford the best model and became which model finishes your specific job for the least money. That is a question you can only answer with your own evals, which is the real work now. Anthropic is also weeks from an IPO after filing confidentially in June, so expect the efficiency drumbeat to get louder, not quieter.

We just launched on X, and we are building it out with the same daily breakdowns you get here.

Today's post covers the Opus 5 launch, the pricing story, and why Anthropic is openly telling you this is not its smartest model. If it is useful, a repost genuinely helps us reach more builders, and following @WorldofAINews means you catch the big drops the moment they happen rather than the next morning.

That is the whole ask. Back to the news.

The stories moving fast beyond today's headlines:

Anthropic launched Claude Opus 5 today, its fourth model in under two months after Mythos 5, Fable 5, and Sonnet 5, available immediately across its platforms and as claude-opus-5 on the API at $5 per million input and $25 per million output, unchanged from Opus 4.8. It is the new default on Claude Max and the strongest model on Claude Pro. Anthropic is explicitly not calling it their smartest model, Fable 5 still holds that, and is instead pitching near-frontier intelligence at half Fable's price, backed by a new low, medium, and high effort dial for trading capability against token spend. The benchmarks are striking: 43.3 percent on Frontier-Bench v0.1 against Opus 4.8's 18.7 and Fable 5's 33.7, three times the next best model on ARC-AGI 3, and a beat on Fable 5's OSWorld 2.0 result at a third of the cost, though every figure is Anthropic's own and the Frontier-Bench runs used Opus 4.8 as a fallback when safety classifiers refused. The behavioral pitch is a model that verifies its own work, demonstrated by a task where it wrote its own computer vision pipeline to read a drawing it could not view, which no competitor solved in five attempts. It stays behind Mythos 5 on cybersecurity by design, carries no data retention requirement for general access, and lands weeks before Anthropic's expected IPO. The broader signal is the one to hold onto: between this, GPT-5.6's efficiency pitch, and China's cheap open weights, the industry has quietly stopped competing on raw capability and started competing on cost per finished task.

Stay building. 🤖