The frontier premium now lasts three weeks

I.Anthropic undercut its own flagship in 21 days

On 22 September Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5/$25 for Opus 5, with cache reads cut from $0.50 to $0.20. Anthropic says the model performs at Claude Fable 5.1 level on most work, costs about 40% less than Opus 5 on typical workloads and generates output more than 30% faster. It is generally available on the Claude Platform, AWS, Google Cloud and Microsoft Azure.

The comparison that matters is not with Opus 5. Fable 5.1, Anthropic's premium tier, launched on 1 September at $10/$50. Three weeks later the same company shipped a model it calls Fable-level at 40% of Fable's per-token price, with Fable-grade safeguards attached: Unite.AI reports three-stage cyber screening that falls back to Opus 4.8 and biology screening that falls back to Opus 5.

The premium tier's exclusive window is now measured in weeks. That changes what a frontier model is for: less a product that earns back its own R&D, more a teacher and a reference price for the tier that actually sells.

II.The lag between flagship and mainstream halved in one cycle

This is the second time this summer. Fable 5 launched on 9 June at $10/$50. Opus 5 followed on 24 July at half the price, and Anthropic said it “approaches Fable-level performance at half the cost”. The Register reported Artificial Analysis putting Opus 5 one index point ahead of Fable 5 at launch. That cycle took 45 days. Fable 5.1 to Opus 5.5 took 21.

Fable 5's first window was shorter still in practice. US export controls imposed on 12 June restricted access for foreign nationals until Anthropic said they were lifted on 30 June, leaving barely three weeks of unrestricted sale before its cheaper sibling arrived.

Buyers had already voted. Ramp data on roughly 70,000 US businesses, reported by the Financial Times on 24 August, showed Fable 5 plateauing at about 11% of spending on Anthropic tools, with Opus 5 overtaking it. Ramp economist Ara Kharazian attributed the shortfall partly to OpenAI's GPT-5.6, priced about 20% below Fable 5. Companies reserved the premium tier for the hardest jobs and routed the rest down.

Set that against how fast capability leaks out of a lab. Epoch AI estimates that the best open-weight models have trailed the closed frontier by about four months on average since January 2026. Anthropic's internal diffusion, from its flagship to its own cheaper tier, now runs roughly five to six times faster than external diffusion to open weights. And Opus 5.5 ships with an anti-distillation safeguard on its preserved reasoning. The lab is distilling its own frontier quickly while making it harder for anyone else to.

III.The honest technical read

Independent measurement supports the headline at maximum effort and blurs it at defaults. Artificial Analysis scores Opus 5.5 at 58 on its Intelligence Index against 53 for Fable 5.1, both at max effort, and measures Opus faster at 92 output tokens per second against 66. At each model's API default, Lindy's reading of the same data has both at 51, with Opus costing about $1.34 a task against Fable's $3.91.

The skeptic's case has three parts. Fable still leads in places: Lindy finds it ahead on GDPval-AA professional work at defaults (1617 against 1576 Elo) and better suited to long unattended agent runs. Token efficiency cuts the other way at max effort, where Opus burns roughly 119,000 output tokens per task against Fable's 78,000, so part of the per-token discount is spent on extra thinking. And Anthropic itself, as Lindy reports, says the gap in Opus's favour is “narrower than these scores suggest”, while Unite.AI quotes the company conceding that benchmark margins are now a less reliable guide at this level.

The 40% figure also needs unpacking. List prices fell 20% against Opus 5. The rest of Anthropic's claim rests on its own measurement of token use on typical workloads, which no independent source we found has replicated. Its showcase example, an HAProxy translation finished in 9.5 hours instead of 12 at 51% lower cost, is a vendor anecdote, not a benchmark.

Safety review is the other caveat. Zvi Mowshowitz's reading of the system card calls Opus 5.5 in effect a smaller version of Fable 5.1, flags weaker refusals when harmful requests are split across conversations, and highlights METR's preliminary estimate of roughly 1.5x AI-driven acceleration in capabilities work. Unite.AI separately notes the system card lists regressions, including readier acceptance of unverifiable authorisation claims. If Fable-level capability now reaches the mass tier in three weeks, each distillate gets about three weeks of scrutiny too.

IV.R&D payback moves down a tier

Prices. Fable-level capability fell 60% in per-token price in 21 days, and roughly 66% in cost per task at defaults on Lindy's numbers. Pricing power at the top of the stack is now bounded by the lab's own release cadence. A buyer who can wait three to six weeks gets the flagship's capability at mass-tier prices.

Capital. Epoch AI's model-lifecycle analysis found GPT-5 earned about $6 billion over a four-month life with roughly $2 billion of gross profit, against about $5 billion of R&D in the preceding months: it did not pay for itself, at gross margins Epoch puts near 30%. A premium window of three weeks makes that arithmetic far harsher. Illustrative estimate, not a sourced figure: take Anthropic's reported $65 billion run rate (Bloomberg via Fortune, 18 August) and apply Ramp's ~11% Fable share to get roughly $7 billion a year of premium-tier revenue; 21 days of that is about $0.4 billion, or around $0.12 billion of gross profit at Epoch's margin. The assumptions are generous — Ramp's US business sample stands in for all revenue, and it ignores that Fable keeps some customers after Opus ships — but the order of magnitude holds. A three-week premium pays for a rounding error of a frontier programme. Payback has to come from volume at the Opus tier, which is exactly where Anthropic is cutting prices.

Labour. At about $1.34 a task for Fable-level work, model cost stops being what limits automation of complex software and knowledge work; review, verification and liability are. Anthropic cites a tester completing a 680,000-line code migration in under a day. That is a vendor anecdote, but it shows where substitution lands first: senior engineering work that was gated by price is now gated by trust.

Market structure. Fast internal distillation is a scale economy. A rival without its own frontier model waits for external diffusion, about four months on Epoch's measure; a lab that distils its own flagship in three weeks takes the volume tier first and protects the source with anti-distillation measures. The durable asset is not the frontier model itself but the speed of turning it into something cheap.

V.Second-order: the public premium tier hollows out

If the public premium tier cannot hold a price, its role shifts to things a cheaper tier cannot offer. Anthropic already sells one: Mythos 5.1, the same model as Fable 5.1 with looser safeguards, restricted to vetted cybersecurity and life-sciences organisations. Opus 5.5 routes most cyber tasks to Opus 4.8 unless the customer joins the Cyber Verification Program. What stays scarce is permission, not intelligence.

Buyers will also learn to wait. If Fable-level capability reaches Opus prices within weeks, procurement teams defer premium commitments, which shortens the premium's revenue window further. Fable 5's 11% plateau suggests that behaviour started before Opus 5.5 arrived.

The third effect falls on safety review. Each fast distillate is a new deployment with its own system card, and the review window shrinks with the release interval. The regressions Anthropic itself disclosed are the cost of that pace made visible.

VI.How it plays out: three scenarios

Probabilities are our analytical judgment, not model forecasts, and they sum to 100%.