Claude Opus 5 is here at half the price of Anthropic’s best model. Here is what actually changed

Anthropic released Claude Opus 5 on Friday, 24 July 2026. It costs $5 per million input tokens and $25 per million output tokens, exactly what the model it replaces cost. Anthropic says it comes close to the intelligence of Fable 5, the company’s most capable public model, at half Fable 5’s price.

Anthropic is not claiming Opus 5 is its smartest model. Fable 5 still holds that position, and a restricted model called Mythos 5 sits above both. Anthropic is claiming Opus 5 is the one you should reach for by default, every day, without thinking about the bill.

First, what these models actually are

If you have only ever used ChatGPT, here is the shape of things.

Anthropic makes Claude, and Claude is not one thing. It is a family of models at different price and capability levels. The ladder runs Haiku 4.5 at $1 per million input tokens, Sonnet 5 at $2 on introductory pricing that expires on 31 August, Opus 5 at $5, and Fable 5 at $10. Above all of them sits Mythos 5, which Anthropic does not sell to the public at all, because it is unusually good at finding and exploiting software vulnerabilities. For the wider picture of who makes what, we ran a guide to the AI tools worth knowing in 2026 last month.

A token is roughly three quarters of a word, so Opus 5 charges about $5 to read a stack of documents the length of eight novels, and $25 to write that much back.

The word you will keep seeing is “agent”. An AI agent is a model given tools and permission to use them in a loop, without a human approving each step. Instead of answering a question, it opens files, runs commands, reads the errors, tries again. A single task can burn hundreds of thousands of tokens, which is why a model that finishes in four attempts at $5 each beats one that finishes in twenty at $2 each. Anthropic’s whole argument for Opus 5 is built on that arithmetic.

What Anthropic says is actually new

The claim is not that Opus 5 knows more. It is that Opus 5 checks its own work and keeps going when the obvious route is blocked. Anthropic gives three examples. All three are its own accounts and none has been independently verified.

In one test, the model was handed a drawing of a machine part and asked to write code rebuilding it as a 3D model in FreeCAD, while being deliberately given no way to look at the drawing. Anthropic says Opus 5 wrote its own computer vision pipeline to pull the geometry out of the raw pixels, then rebuilt the part, and repeated the feat. No competing model solved it in five attempts.

In another, the model was given a real bug in a widely used open-source package manager. Anthropic says Opus 5 traced the root cause and fixed an edge case the community’s own patch had missed, while a rival model fixed the visible symptom and declared the bug resolved.

In the third, an engineer at a trading firm built a market data feed for a new exchange in one session. Earlier models could not finish the job even with detailed plans. With no live feed to test against, Anthropic says the model built its own test harness to check it was parsing the exchange’s data correctly.

Anthropic also published testimonials from 22 early-access customers. These were selected by Anthropic and should be read that way, but they are on the record with names attached. Wade Foster, chief executive of Zapier, says Opus 5 topped his company’s own AutomationBench leaderboard and scored 100% on a customer-churn task no previous model passed. Ben Kus, chief technology officer at Box, reports Opus 5 beating Opus 4.8 by 8% overall, with an 11% gain on data analysis and 17% on due diligence.

The most useful entry is not a number. Marquis Wang, a principal AI engineer, describes the model rejecting a design he proposed and holding its position when he insisted. It narrowed its objection to a single question and offered a compromise. Anthropic is selling judgement, and that is what a customer chose to describe.

What the numbers say

Anthropic published results against Fable 5, its own previous Opus 4.8, and OpenAI’s GPT-5.6 Sol. Several of the benchmarks are built by outside organisations, but the runs are Anthropic’s own rather than independent, and that caveat applies to every figure below.

One footnote on the announcement deserves more attention than a footnote usually gets. On the Frontier-Bench chart, Anthropic says Opus 4.8 stood in as a fallback whenever a safety classifier refused a request from Opus 5 or Fable 5. Some portion of both scores therefore reflects work done by an older model. Anthropic does not say how large that portion is.

On Frontier-Bench v0.1, a test of coding work done through a terminal, Opus 5 scored 43.3%. Fable 5 scored 33.7% and GPT-5.6 Sol scored 34.4%. On OSWorld 2.0, which measures whether a model can actually operate a computer by clicking and typing the way a person does, Opus 5 scored 70.6% against Fable 5’s 66.1%. On AutomationBench, which checks whether a model can finish an ordinary business task from start to end, Opus 5 passed 26.0% against 18.1% for GPT-5.6 Sol. On BrowseComp, which tests whether a model can find hard-to-locate information on the open web, it scored 90.8%.

The knowledge-work result is the one Anthropic leads with, and the number looks odd until you know how it is made. GDPval is a dataset OpenAI built from real professional deliverables across 44 occupations in the nine largest sectors of the US economy, written by people averaging 14 years in their fields. Independent firm Artificial Analysis runs its own version, has judges compare outputs blind, then converts the win record into a chess-style rating. Opus 5 scored 1,861, against 1,747 for Fable 5, 1,736 for GPT-5.6 Sol and 1,593 for Opus 4.8. Anthropic is using a benchmark built on a rival’s dataset to claim the lead in ordinary office work.

The largest gap appears on ARC-AGI-3. That benchmark drops a model into a small game world with no instructions and no stated goal. It has to work out the rules by playing. When the ARC Prize Foundation launched it in March 2026, every frontier model scored under 1%, while human testers solved all of it. Opus 5 scored 30.2%. GPT-5.6 Sol scored 7.8%. Opus 4.8 scored 1.5%.

Where it loses

Anthropic lost four of these tests and published them anyway. On DeepSWE v1.1, another coding evaluation, GPT-5.6 Sol scored 72.7% against Opus 5’s 68.8%. On Humanity’s Last Exam without tools, Fable 5 edged it, 56.5% to 56.3%. On a held-out legal benchmark Fable 5 won, 13.3% to 11.7%. On HealthBench Professional the restricted Mythos 5 led at 66.0% against Opus 5’s 59.8%.

On science, Anthropic reports Opus 5 beating Opus 4.8 on every life sciences evaluation it ran, with the biggest gains in organic chemistry. Working out a molecule’s structure from spectroscopy data improved by 10.2 percentage points, and predicting how changes in a protein’s sequence affect its function by 7.7.

One more figure cuts against the marketing instinct. Anthropic’s automated behavioural audit scores models from 1 to 10 on misaligned behaviour, and lower is better. Opus 5 scored 2.30, its best to date. Sonnet 5, released last month, scored 3.35, the worst of the four models tested.

Security, access tiers, and what happens to your data

Anthropic separates finding a software flaw from building a working attack out of it. On an evaluation using OSS-Fuzz, Opus 5 identified vulnerabilities in 79.4% of challenges, almost matching the restricted Mythos 5 at 80.0%. On turning those flaws into exploits, Mythos 5 succeeded on 13 challenges and Opus 5 on four. That gap is why Opus 5 ships to everyone and Mythos 5 does not.

Anthropic says Opus 5’s filters block binary-based vulnerability scanning, penetration testing and exploit generation, and should trigger about 85% less often than Fable 5’s. When they trigger inside Claude.ai, Claude Code or Claude Cowork, the request quietly falls back to the older Opus 4.8. Access is therefore two-tier: enterprises and researchers already admitted to Anthropic’s Cyber Verification Programme get a version with fewer restrictions straight away. Blocked biology requests have also moved. They now route from Fable 5 to Opus 5 rather than Opus 4.8, which makes Opus 5 the most capable model the public can use for scientific research.

For any business weighing this, one line in the announcement matters more than the benchmarks. Opus 5 carries no data retention requirement for general access, so prompts are not held for 30 days the way Fable 5’s are. If you handle client records, medical data or anything covered by Kenya’s Data Protection Act, that is the sentence to take to your compliance officer.

What it costs in Kenya

Claude is available here, billed in dollars, and Kenya charges 16% VAT on digital services from non-resident companies under the VAT (Electronic, Internet and Digital Marketplace Supply) Regulations. There is no minimum threshold, so the tax applies from the first shilling.

At roughly KES 129 to the dollar, Claude Pro at $20 lands near KES 3,000 a month once VAT is added. The $100 Max plan, where Opus 5 is now the default and Fable 5 is included, works out closer to KES 15,000. The $200 tier is about KES 30,000. We laid out this arithmetic in full when Anthropic started pricing Claude in rupees for India while Kenya kept paying in dollars earlier this month. KNBS puts average formal-sector earnings at about KES 82,000 a month before deductions. Claude Pro is around 3.6% of that. The cheaper Max plan is around 18%.

Opus 5 becoming the Max default is the most useful part of today’s news for subscribers already paying that. For everyone else it changes nothing about the price of entry. To see what these models do in ordinary use, we set up Glovo’s shopping assistant inside Claude and ordered snacks in Kasarani a fortnight ago.

What to watch

Opus 5 is available today on Anthropic’s platforms and as claude-opus-5 on its API, with a Fast mode at roughly 2.5 times the speed for double the price. It is the default on Max and the strongest option on Pro. Two smaller changes ship in beta alongside it: developers can now swap which tools Claude has access to mid-conversation without losing the cached prompt, and API requests flagged by safety filters can be routed automatically to another model instead of being refused.

Two things will decide whether the launch works. The first is whether the efficiency claims hold once real workloads hit the model rather than benchmark suites, since every figure above is Anthropic’s own and no independent testing exists yet. Watch for scores from Artificial Analysis and the coding leaderboards over the next fortnight.

The second is how people react to the fallback system. Requests that trip a safety filter get answered by an older, weaker model. When Fable 5 returned from a government ban in July, subscribers noticed that happening and complained loudly.

The price is the other half of this story, and it was set by pressure from outside Anthropic. OpenAI shipped a rival flagship two weeks ago, China released the largest open-weight model ever built eight days ago, and Google has missed its own deadline by two months. We explain why Opus 5 costs what it costs, and who forced it in a companion piece.