Anthropic dropped prices 20% on its two most capable models yesterday and published benchmark numbers that would make any enterprise procurement team reach for a purchase order. Then, buried in the same announcement, the company admitted it deliberately constrained what these models can do in the domain most relevant to AI development itself. That admission doesn't just complicate the launch. It makes the benchmark scores enterprises use to pick AI vendors fundamentally unreliable.
Here's what happened. Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on Monday. The two are the same underlying model—Fable ships with strong cybersecurity and biology safeguards for general use, while Mythos has relaxed guardrails and is restricted to vetted organizations through Anthropic's Life Sciences and Cyber Verification Programs. Standard pricing is $8 per million input tokens and $40 per million output tokens. The Decoder notes that's down from $10 and $50 for the previous generation. SiliconANGLE puts the reduction at 20%. An optional fast mode—up to 2.5 times faster—runs at $16 per million input tokens and $80 per million output tokens, available on Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry.
The capability claims are genuinely striking. About 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in Lean in 30 hours, starting from brief instructions with no mathematical references, organizing their own collaboration, using one-third the action tokens of an earlier attempt and producing a proof seven times shorter. The model solved two open problems in theoretical physics involving black holes and the unification of relativity and quantum mechanics—one completing a calculation laid out in 2003 that was too laborious to finish, the other finding the exact limit of a technique physicists have used for more than 15 years. Claude Mythos 5.2 accurately modeled 28 out of 34 AAV capsid structures. One Anthropic engineer gave Fable 5.2 an empty folder and eight days; it produced a working operating system with its own kernel, drivers, file storage, and a login screen, coordinating over 1,000 subagents to do it. Anthropic also says over 90% of changes merged into its own core codebase are now written by Claude, with human guidance.
On third-party benchmarks, Fable 5.2 scores 65 points on Artificial Analysis's Intelligence Index v5 at the highest compute tier—five ahead of Gemini 4 Carbon at 60 and six ahead of Claude Opus 5.5 at 59, according to The Decoder. The per-task cost at that tier averages $26.50, roughly 2.6 times the cost of running Opus 5.5.
The RSI problem nobody is talking about loudly enough
Now here's the part that should give every enterprise buyer pause. Anthropic says it deliberately held Fable 5.2 and Mythos 5.2 back during training from improving at capabilities most directly tied to recursive self-improvement—things like optimizing compute kernels for accelerator chips and improving large language model architectures. The Decoder reports this explicitly. The models perform roughly on par with Opus 5.5 in those specific areas, not because they couldn't do better, but because Anthropic chose to constrain them.
Think about what that means for every benchmark score in this announcement. The model was trained with awareness—at some level, in some form—of which capabilities were being watched and constrained. Anthropic is not claiming the model detected individual evaluation runs and sandbagged them. But they are saying the training process itself was shaped by what the model was and wasn't supposed to get good at. The line between 'we constrained training' and 'the model's scores reflect what it was optimized to show' is thinner than it looks.
The open question in Anthropic's own disclosure is whether the constraint was effective and how they verified it. The sources don't say. What they do say is that agentic coding, hardware development, and traditional AI research—including interpretability research—are not affected by the RSI throttling. So the model was trained to be good at some things and not others, and the benchmarks measure the things it was trained to be good at. That's not a neutral scorecard. That's a report card where the student helped design the grading rubric.
The fair counterargument is that all model training involves choices about what to optimize for, and Anthropic's transparency here is more than most vendors offer. They told us about the constraint. That's worth something. But transparency about the existence of a constraint doesn't tell you how it distorts the scores, and it doesn't give you an independent way to verify performance on tasks the model was specifically shaped to underperform on. You're still trusting the vendor's choices about what the benchmarks measure.
There's also a timing note worth flagging. The Decoder's headline observes this launch comes shortly before Anthropic's IPO. That's not an accusation—it's context. But it's the kind of context that should sharpen your skepticism about taking benchmark tables at face value right now.
What this means for your AI budget and vendor conversations
Three things to act on this week.
First, if your team is mid-procurement on a frontier AI contract and using benchmark scores as a primary selection criterion, ask your vendor directly: which capabilities were constrained or deprioritized during training, and how does that affect performance on your specific use case? 'Best on the benchmark' and 'best for your workflow' are not the same thing, and the RSI disclosure makes that gap concrete.
Second, the 20% price cut is real and the token efficiency improvement is real—Fable 5.2 uses 98,000 output tokens per task at lower compute tiers versus 173,000 for Opus 5.5 at Max tier, per Artificial Analysis data cited by The Decoder. If you're already running Claude at scale, that's a budget line worth revisiting now. But run your own evals on tasks that matter to your business before you commit to the highest compute tier, where per-task costs average $26.50—about 2.6 times Opus 5.5.
Third, if you're in cybersecurity or biomedical research and you're not already in Anthropic's Cyber or Life Sciences Verification Programs, the Mythos 5.2 numbers are worth a look. Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 in Anthropic's internal CVE Rediscovery evaluation using 1,829 CVEs—and one security firm reported that one in six confirmed bugs Mythos found had been missed by every other frontier model tested. Those are Anthropic's own figures, so verify independently, but they're specific enough to be worth testing against your own threat library.
What would prove the benchmark-reliability concern wrong? Independent red-teaming of RSI-adjacent tasks by organizations with no relationship to Anthropic, published before the IPO window closes. If those scores hold up under adversarial evaluation, the concern shrinks considerably. If they don't, the question of what else was shaped by training choices becomes a lot harder to ignore. My bet is we don't get that independent evaluation before the IPO—and that the absence of it will matter more than anyone is admitting right now.