The pricing change is where this is worth explaining properly rather than just reporting the number.

An API, the connection mentioned above, is usually billed by how much text is sent to the AI and how much comes back, measured in small chunks of text called tokens.

One part of that bill is a "cache read", which is what happens when the AI reuses text it has already processed rather than reading it fresh each time. Think of it like a photocopy: reusing a copy that already exists is cheaper than printing a brand new one every time.

Anthropic said it kept its standard pricing the same, but cut the price of those cache reads by 75%, from $1 down to $0.25 per million tokens.

Anthropic said that brings typical use "25% less than Fable 5" overall, and up to roughly 45% cheaper for multi-step, automated tasks that keep reusing the same background information, since those lean hardest on cached reads.

In practice, that saving mostly lands with the businesses and developers paying for Claude by the token, not with someone on a personal Claude subscription chatting normally. If it gets passed on, it would most likely show up as a slightly cheaper or more capable AI-powered tool somewhere else, not as a bill reduction anyone sees directly.

Anthropic also shared performance numbers to show the new models are not just cheaper, but more capable too.

These come from tests researchers use to compare different AI models against the same tasks, similar to how a driving test score lets two different drivers be compared fairly.

Anthropic reported Fable 5.1 scoring 52.6% on one such test, called Terminal-Bench-Science, which checks how well an AI can carry out real scientific and technical tasks on its own, up from 24.7% for the previous Fable 5.

That is a genuine improvement, but worth reading carefully. Even at 52.6%, the model still gets these tasks wrong roughly half the time, so it is not yet something a researcher could rely on unsupervised.

Mythos 5.1 scored 60.9% on a separate coding-focused test called Terminal-Bench 4.0, Anthropic said. Anthropic also reported it correctly designed usable proteins, the building blocks used in a lot of biology and drug research, close to half the time across 12 test cases, against a typical success rate of one in ten to one in seven for that kind of task. If that pattern holds up outside Anthropic's own testing, it points to a meaningfully faster starting point for early-stage biology research, not a finished result on its own.

Two days later, on 3 September 2026, OpenAI released its own new model, GPT-6 Astra, which it called "the world's most intelligent and aligned model", according to CNBC's coverage of the announcement.

OpenAI said, according to CNBC, that Astra sets a new state of the art across computer use, web browsing, software engineering, cybersecurity, science, and general professional work, and that it can create documents, spreadsheets, and presentations directly, rather than just describing how to make them.

OpenAI's own test scores, as reported by CNBC, are steep. It scored 98% on a test of advanced maths problems called FrontierMath Tier 4, and 99.9% on a general reasoning test called ARC-AGI-3, a set of puzzles specifically designed to be easy for a person and historically difficult for AI.

It also scored 100% on a cybersecurity test called ExploitBench, OpenAI said, which measures how well a model can find exploitable security flaws on its own.

That last score cuts both ways. The same skill that helps a security team find and fix a weakness before an attacker does could, in the wrong hands, help someone find that same weakness in order to exploit it.

OpenAI, via CNBC, also said Astra has already helped make progress on some long-standing open maths problems, though that is OpenAI's own claim about its own model, not something independently checked yet.

The rollout started 3 September 2026 with early access for OpenAI's largest business customers, through what it calls "Daybreak" access, CNBC reported.

OpenAI said wider access for ChatGPT Plus, Pro, Business, and Enterprise users, plus its API and Amazon Web Services, would follow "in the coming days".

Al Jazeera's coverage noted the release landed amid rising scrutiny and safety concerns, more than a typical model release tends to draw, which lines up with the exploit-finding score above: a model this capable at finding security flaws draws closer attention to how it might be misused, not just how well it performs.

Why it actually matters

For most people already paying for ChatGPT or Claude to chat normally, nothing changes today.

Astra is rolling out gradually to OpenAI's largest customers first, so anyone on a personal ChatGPT Plus or Pro plan will likely get it in the coming days rather than immediately, and it will most likely arrive quietly rather than get announced with fanfare.

The genuinely useful change, the Claude pricing cut, mostly matters to a different group of people: anyone building or paying for a tool, app, or automated service that runs on the Claude API behind the scenes, rather than someone typing into Claude's chat window directly.

For someone who just opens ChatGPT or Claude to ask questions day to day, the honest answer is that this week's releases change very little right now. The test scores climb and the models get sharper at technical work, but for typical everyday use like drafting an email, checking a fact, or planning a meal, the difference is unlikely to feel dramatic.

Where it is worth paying attention is the direction both companies are moving in: AI models getting measurably better at technical, multi-step tasks like coding, research and security work, while getting cheaper to run those tasks at scale. That shift is more likely to show up as better AI-powered tools over time than as anything a chat user notices directly this week.

What's still unclear

Neither company's own announcement makes clear whether these updates change the everyday Claude or ChatGPT chat apps that most individual readers actually use, as opposed to the developer and business-facing products listed above.

OpenAI's test scores for Astra are OpenAI's own, run on OpenAI's own terms, and have not yet been independently tested by outside researchers.

The specifics behind the "rising scrutiny and safety concerns" Al Jazeera flagged around Astra's release have also not been spelled out in detail anywhere reviewed for this piece, only that the release drew more attention on that front than usual.

It is also not clear whether Mythos-level access will ever open beyond the current small group of vetted cybersecurity and life-sciences organisations, or stay permanently restricted.