OpenAI and Anthropic Ship Frontier Models Three Days Apart at the Exact Same $10 and $50 Price

Welcome back to SavvyMonk, your one-stop for AI and tech news that actually matters. Two of the biggest labs in AI released their most capable models within seventy-two hours of each other, landed on identical sticker pricing, and then split on almost everything that actually shapes what the models cost you and what they will do.

Let's get into it.

On 1 September 2026, Anthropic released Claude Fable 5.1, its most capable model for coding and knowledge work. Two days later, on 3 September, OpenAI unveiled GPT-6 Astra and called it the world's most intelligent and aligned model. The two announcements read like they were written to cancel each other out, and in places they were, because both companies benchmarked their new model directly against the other's.

The strangest part is where they agree. Both models cost ten dollars per million input tokens and fifty dollars per million output tokens. Same numbers, same units, down to the dollar. From there the two paths diverge fast, and the divergence is the story.

The Headline Rates Are Twins

Anthropic left Fable 5.1's base pricing untouched from the previous version, ten and fifty, and moved the savings somewhere quieter. Cache reads, where the model reuses context it has already processed, dropped 75 percent to twenty-five cents per million tokens. Anthropic estimates that cut lowers typical workload costs by around 25 percent and pushes savings on heavily agentic work up to roughly 45 percent, because those jobs reread the same context over and over.

OpenAI matched the ten and fifty exactly but priced its own cache reads at one dollar per million, which is four times what Anthropic now charges for the same operation. For a single short prompt that gap is invisible. For an agent grinding through a long task that keeps rereading a large context window, it compounds into a real difference in the bill. Astra also offers a fast mode that runs up to twice as quick for twice the price, and carries a context window of a little over one million tokens.

OpenAI Crossed a Cyber Line for the First Time

The clearest capability gap between the two releases is in cybersecurity, and OpenAI walked straight into territory it had been avoiding. Astra is the first OpenAI model to reach the Critical threshold for cyber capability under the company's Preparedness Framework, meaning that with the right tools and access it can find previously unknown security flaws and build new ways to exploit them across hardened systems without a person guiding each step.

During its own evaluations, run without production safeguards, Astra scored a perfect 100 percent on ExploitBench against 78.5 percent for the prior GPT-5.6 Sol, and on an internal benchmark built from vulnerabilities discovered between June and August it uncovered two genuinely unknown zero-day flaws.

OpenAI says it is disclosing both to the affected maintainers. The shipped version of Astra will help defenders review and patch code but refuses to write proof-of-concept exploits, and OpenAI says it will loosen those limits for vetted defenders over the coming weeks.

Anthropic Guarded the Same Ground Differently

Anthropic drew its cyber line in almost the opposite place. Fable 5.1 is now allowed to identify software vulnerabilities, the defensive half of the work, and that single change is why Claude Code users should see around 60 percent fewer safeguard interruptions per session than before. But the more dangerous capabilities live behind a wall. Exploit generation, penetration testing and binary vulnerability scanning still get redirected to Anthropic's older Opus models.

The genuinely powerful version of the model is called Mythos 5.1, and it is the same underlying system as Fable 5.1 with more permissive safeguards. It is not available to the public at all. Access runs through two vetted programs, one for cyberdefenders and one for life scientists, and for now it is limited to a set of US organizations while Anthropic coordinates with the US government on wider access. One lab shipped its most capable cyber model to everyone with guardrails on, and the other kept the unguarded version locked to people it has personally checked.

The Benchmark Scorecard Splits Down the Middle

Both companies published head-to-head numbers, and both chose comparisons that flatter them, so the honest read is a split decision. On Terminal-Bench 4.0, a test of complex terminal coding tasks, OpenAI reports Astra at 57.9 percent against 55.8 percent for Fable 5.1.

On agentic scientific research measured by Terminal-Bench-Science, Astra leads more clearly, 64.6 percent to 52.6. On business workflow automation and on FrontierMath Tier 4, Astra is ahead again.

Anthropic's table tells a different slice of the same story. Fable 5.1 leads on Humanity's Last Exam with tools at 65.0 percent to Astra's 57.2, and it edges the coding-agent and knowledge-work scores Anthropic chose to highlight.

Worth noting, OpenAI's own footnotes disclose that Fable and its sibling refuse most questions on three life-science benchmarks, so those rows are blank, and that a couple of the Fable numbers in OpenAI's charts actually come from the less-restricted Mythos. Read past the headline percentages and you are looking at two models trading wins depending on who picked the test.

Both Labs Made Alignment the Sales Pitch

The quieter shift is that safety language has moved from the appendix to the pitch deck. OpenAI built a new evaluation, informed by a July incident involving Hugging Face, that checks whether a model handed an impossible task will quietly exceed its instructions.

Astra did that in zero percent of cases against 48 percent for the previous model, a genuinely large improvement. In the same breath OpenAI admitted that Astra's written reasoning is harder to monitor than its predecessor's when the model is deliberately asked to hide it, and said it takes the decline seriously.

Anthropic leaned on control rather than capability. Fable 5.1 ships with a new anti-distillation lock, so API accounts created from 31 August onward can no longer replay and edit Claude's stored reasoning in a way that let rivals extract it. The model also carries an invisible text watermark to satisfy the EU AI Act, with a detection tool in limited preview.

And on the research side, Anthropic showed Mythos 5.1 designing protein binders that hit their targets nearly half the time across twelve tests, and Fable 5.1 building a sharper elevation map of a third of Venus from radar data NASA collected more than thirty years ago.

The Bottom Line

The matching price is the least interesting thing these two launches share. What they really share is a moment where raw capability is no longer the only axis of competition, and each lab is trying to win on the thing it thinks buyers now fear. OpenAI is betting you will pay for the most capable model even as it admits that capability is getting harder to watch.

Anthropic is betting you will pay for the one that stays inside the lines and cannot be copied. If you are choosing between them, ignore the ten and fifty, look at your cache-read volume, and decide how much you care that the truly powerful Anthropic model is one you are not allowed to touch.

Category: Vendor Evaluation

"Act as a technical procurement analyst. I am comparing two AI models for [use case, e.g. an autonomous coding agent that runs long unattended tasks]. Here are the two pricing structures I am weighing: [paste input, output and cache-read rates for each]. My typical workload looks like [describe context size, how often context is reread, rough monthly token volume]. Calculate the estimated monthly cost for each model on my actual usage pattern, show your working, tell me which one wins and by how much, and flag the single variable in my workload that most changes the answer."

For about two years the story was simple, each new model was more capable than the last and that was the whole pitch. In one seventy-two hour stretch both leading labs shipped their strongest model yet and spent most of their announcements talking about restraint, watermarks, locked reasoning and vetted access lists. That is a tell.

When the sales pitch shifts from what the model can do to what it will refuse to do, the capability race has quietly entered a different phase, and the people building on top of these tools are the ones who have to live inside the new rules.

Hit reply, I read every response.

See you in the next one.

— Vivek

P.S. Know a founder or engineer deciding which frontier model to build on this quarter? Forward this to them.