Anthropic's Haiku 5.5 Drops Sub-Frontier Pricing 90% — and Reorders the $518B Compute Stack

Anthropic shipped Claude Haiku 5.5 on October 7, 2026 — the third leg of the Claude 5.5 family after Opus 5.5 on September 22 and Sonnet 5.5 on September 28. The sub-frontier tier is the smallest model in the lineup but carries the heaviest lift: per Anthropic's release notes, roughly 90% of requests to the previous Haiku model landed at prompts under 100,000 tokens — exactly the band where Haiku 5.5 prices 90% below its predecessor.

The headline sticker is $0.10 per million input tokens and $0.50 per million output tokens — the same rate OpenAI set on GPT-6 Luna two weeks earlier. Anthropic's own page positions the model against GPT-6 Luna on six benchmarks and wins five of them, most decisively on agentic coding (Terminal-Bench 4.0: 39.2% vs 70.6%, where Luna still leads) and computer use (OSWorld 2.1: 72.4% vs83.9%).

The price cliff matters more than the headline rate. Above 100,000-token prompts, Haiku 5.5 jumps fivefold to $0.50 / $2.50 — at which point Sonnet 5.5 at $2 / $10 with stronger intelligence becomes the rational routing choice for long-context jobs. Anthropic is not just undercutting OpenAI; it is mechanically nudging high-context traffic from the cheap tier to the mid tier, protecting blended average selling price per query even as the entry tier gets louder.

The Anthropic release page leads the marketing with per-token pricing. The third-party measurements lead with per-task economics. Artificial Analysis's Intelligence Index testing shows Haiku 5.5 at max effort burns roughly 162,000 output tokens per task against GPT-6 Luna's ~50,000 — meaning Haiku 5.5 actually costs roughly 3× more per completed job at max effort despite identical sticker rates. Customers who route on per-task cost, not per-token sticker, will land on Luna unless Haiku's intelligence lift justifies the burn.

The price ladder now spans 200× from bottom to top — Haiku 5.5 at $0.10 input to Opus 5.5 at $4 input — and each tier has been cut, repositioned, or both in the run-up to the IPO. Opus 5.5 trades 20% off sticker plus 60% off cache reads and now matches GPT-6 Astra on Terminal-Bench at about 40% of the cost per task. Sonnet 5.5 kept its $2 / $10 sticker but re-architected to use ~30% fewer tokens per task on hard coding workloads, which is what drives the \"30% cheaper per task\" claim.

For the IPO math, the stack is more important than any single tier. If cheap inference is the volume flywheel that fills Anthropic's $518 billion in committed cloud and infrastructure spend, Haiku 5.5 is the tier that converts the flywheel's rotations into revenue. If mid- and high-tier reasoning contracts dominate instead, the Haiku tier becomes a marketing line item. The October 7 launch is Anthropic betting on the first outcome — and pricing it so that even with GPT-6 Luna matching on sticker, the volume stays inside Anthropic's APIs.

Cheap inference does not materialize out of thin air — it runs on the $518 billion in committed cloud, compute, and infrastructure obligations Anthropic disclosed in its IPO prospectus. Per the Reuters review of that prospectus, the commitments break down as roughly $111 billion to Alphabet (Google TPU via Broadcom), $110 billion to Amazon (AWS Trainium + NVIDIA in AWS), $31.4 billion to Microsoft (Azure compute), and $161.2 billion to Broadcom (custom silicon for Google TPU and Anthropic-adjacent programs). About 80% of those obligations are binding and non-cancelable. For every $1 Anthropic earned in 2025, it has committed roughly $113 to future infrastructure — the leverage ratio that makes the Haiku 5.5 volume bet unavoidable.

Short-term (days to quarters), the volume reads off Haiku 5.5's customer migration. Anthropic cites Asana reporting \"over 30% reduction in latency\" and \"up to 2.5× faster inference per agent turn\" versus Haiku 4.5, and HubSpot scoring 92.8% averaged over three runs against its prior default. Those data points are the early signal that sub-frontier workloads are now viable at sub-second latency — the condition for the cheap-inference volume bet to work.

Medium-term (next two to four quarters), the watchpoints are utilization at the hyperscaler three and CoreWeave's revenue conversion. NVIDIA's Q2 FY2027 data-center revenue of $89 billion (up 117% year-on-year, reported Aug 26, 2026) is the upstream tide that lifts all of these names. CoreWeave, with negative trailing free cash flow and 11.5× trailing net-debt-to-EBITDA per the FY2025 10-K, is the most leveraged single name to Anthropic's volume bet — its FY2025 revenue of $5.13 billion and 1.3 GW of contracted power capacity are the numbers that move first if Haiku 5.5 volume holds.

Long-term (one to three years), the structural question is whether the $0.10 / $0.50 sub-frontier floor collapses further or stabilizes. Gemini 3.8 Flash's introductory $0.75 / $3.75 rate expires by end-2026; if Google re-prices under Haiku/Luna, the sub-frontier tier enters a deeper deflation round that compresses every participant's per-token economics. The flip side is that Anthropic's reported $518B obligations, $20.28B cash position, and Anthropic-AWS $100B+ commitment create a captive-volume moat that makes OpenAI's matching price cut hurt more than it helps.

Anthropic's reported $2 trillion IPO valuation, on the economics visible in the IPO prospectus reviewed by Reuters, cannot be supported by high-margin Opus and Sonnet contracts alone — those drive revenue concentration, not the utilization that fills the $518 billion in cloud obligations. The bet Anthropic is making with the October 7 launch is that Haiku 5.5's volume tier generates the queries that justify the build-out, and that tiered pricing at the 100,000-token cliff protects blended revenue per query even as the entry sticker drops 90%.

OpenAI matching the cut on the same day as GPT-6 Luna was not an accident. With both frontier labs pricing the sub-frontier tier at $0.10 / $0.50 and one of them (Anthropic) running ~3× the tokens per task to win on intelligence, the per-token price war is effectively over and the contest has moved to per-task economics and customer lock-in. Hyperscalers and silicon suppliers win either way — the only listed-company loser is any AI app whose gross margin depends on a per-token spread that no longer exists at the sub-frontier tier.