Sonnet 5.5 priced like Sonnet 5, benchmarked like Opus 5.5 — Anthropic's mid-tier reset ahead of the $2T IPO

Anthropic shipped Claude Sonnet 5.5 on September 28, 2026 at $2 per million input tokens, $10 per million output tokens and $0.20 per million cache reads — verbatim the rate card Sonnet 5 has carried since its June 30 launch. Cache writes sit at $2.50 per million. No introductory discount, no trailing window: the introductory rate Sonnet 5 carried through August 31 was made permanent in August 2026, so Sonnet 5.5 inherits a price floor, not a price cut.

What changed is what the same dollars buy. Sonnet 5.5 generates output roughly 30% faster than Sonnet 5 and, per Anthropic, runs many tasks for up to 30% less than Sonnet 5 — even though the per-token price is identical. The mechanism is a longer-thinking, higher-quality response: more tokens per task, fewer tasks per workflow. It is a deliberate trade of tokens for capability, not a price war.

On Anthropic's own published benchmarks, Sonnet 5.5 lands within rounding distance of Opus 5.5 on enterprise-relevant workloads — at half the per-token price. On Terminal-Bench 4.0 (agentic coding), Sonnet 5.5 scores 70.6% versus Opus 5.5's 66.4% and Sonnet 5's 10.3%. On GDPval-AA v2.1 (knowledge work), Sonnet 5.5 scores 1,844 versus Opus 5.5's 1,846 — a 0.1% gap that is well inside measurement noise. On AA-Briefcase v1.1, Sonnet 5.5 hits 1,811 versus Opus 5.5's 1,822.

The Artificial Analysis Intelligence Index independently puts Sonnet 5.5 at #2 with a score of 56 — two points behind Opus 5.5 and tied with GPT-6 Astra on Terminal-Bench 4.0 at 64%. The same evaluator notes that Sonnet 5.5 uses roughly 60% more output tokens per task than Opus 5.5 (193k versus ~120k), which is why it sits off the cost-vs-intelligence Pareto frontier at default effort despite the cheap per-token rate. The pricing lever works in the customer's favor only when Sonnet 5.5's extra capability lets them skip a workflow step they used to pay Opus for.

Six days before Sonnet 5.5 landed, OpenAI shipped GPT-6 Sol and GPT-6 Luna on September 22, 2026 — the same day Anthropic shipped Opus 5.5. Sol is priced at $2/$10 per million tokens (input/output), exactly matching Sonnet 5.5; Luna is priced at $0.10/$0.50, 50% below GPT-5.6 Luna's promotional rate. The two flagship launches landed within hours of each other and converged on the same mid-tier price.

Gemini 3 Pro sits at the same $2 input tier. The mid-tier of the frontier model market is now $2 in and $10 to $12 out across all three Western frontier labs — differentiation has moved from sticker price to capability per token, and Anthropic is the one with a model that just printed near-Opus scores at the shared rate. OpenAI's response is structural: Luna at 5% of Sol's input price drags the floor of the budget tier down with it, forcing Anthropic's expected Haiku 5.5 to land below the $2 line or cede the high-volume, low-stakes workload to Luna.

Anthropic filed a confidential S-1 on June 1, 2026 at a $965 billion private valuation, told prospective investors Q2 2026 revenue exceeded $11.5 billion, and shifted the listing window toward mid-October, with bankers and investors floating a $2 trillion target. The model cadence since has been methodical: Opus 5 / Sonnet 5 on June 30 and July 24, Fable 5.1 / Mythos 5.1 on September 1, Opus 5.5 on September 22, Sonnet 5.5 on September 28 — each on a roughly three-to-six-week cadence and each with a lower per-token price than its predecessor.

The math that matters for the S-1 is revenue divided by token price, not revenue alone. Anthropic's ARR scaled from roughly $9 billion at year-end 2025 to about $47 billion in May2026 to ~$65 billion by July — more than sevenfold in seven months — on a per-token price that has fallen at every step. Sonnet 5.5 at $2/$10 with near-Opus capability is the cleanest expression of that thesis: customers burn more tokens on harder problems, revenue compounds, and gross margins hold because Anthropic's inference cost is collapsing faster than its list price.

Anthropic has now signed compute commitments worth more than $500 billion across more than a dozen counterparties through October 2025–September 2026, with roughly 14.8 gigawatts of capacity lined up over the decade. The biggest single line is the Amazon deal: more than $100 billion to AWS over ten years, securing up to 5 gigawatts of capacity spanning Graviton and Trainium2 through Trainium4 silicon, with nearly 1 gigawatt of Trainium2 + Trainium3 capacity coming online by the end of 2026. Anthropic says it already runs over one million Trainium2 chips to train and serve Claude today.

Microsoft holds a $30 billion Azure compute commitment plus up to one additional gigawatt; Claude has been available in Microsoft Foundry since June 30, 2026. Alphabet's Google Cloud expansion, first announced in October 2025, brings well over a gigawatt of TPU capacity online through 2026 and adds tens of billions of dollars to Anthropic's compute bill. AWS revenue from Anthropic is on track to climb from $1.28 billion in 2025 to nearly $3 billion in 2026 and $5.6 billion in 2027, per analyst estimates.

The wildcard is NVIDIA. NVIDIA supplies the chips, holds the lease on the underlying data center, and is in talks to invest in Anthropic's IPO — a triple role in the same transaction that drew circular-financing criticism after the $35 billion round with Lambda and Hut 8 closed in early September 2026. NVIDIA's Q2 FY27 disclosures show $279 billion in supply obligations and $108.5 billion in guarantees to AI cloud and data-center partners; AI-lab demand is on track to be roughly a quarter of next year's business, per CFO Colette Kress. Every Anthropic price cut is, at the infrastructure level, an NVIDIA volume event — Anthropic's cheaper tokens still burn on NVIDIA silicon through Lambda, and AWS' Trainium is additive rather than substitutive for the most demanding training runs.

Sonnet 5 is already the default Anthropic model behind Claude Code, and it is generally available for GitHub Copilot across Pro, Pro+, Max, Business and Enterprise tiers since June 30, 2026. Sonnet 5.5 inherits that distribution on day one; GitHub Copilot, Cursor and the various Claude Code-powered IDEs all sit on Anthropic's API at the new rate card. Cursor pricing is unchanged at $40/seat on Teams and $120 on Enterprise; Copilot Enterprise at $39/seat is the most operationally simple. The bet from those platforms is that a 30% cut in cost per task lets them absorb more usage per seat without raising list prices — which protects subscriber churn while lifting gross margin.

The bigger enterprise slot opened on August 26, 2026, when Salesforce and Anthropic announced Claudeforce — putting Salesforce CRM data inside Claude via governed permissions, with a broader open beta planned for September 2026 and the Salesforce-in-Claude rollout already live to pilot customers. Deloitte is already expanding its collaboration around the same stack. With Sonnet 5.5 at near-Opus quality for agentic coding and knowledge work, Salesforce gets a Claude-powered reasoning layer for agentic front-office workflows at a per-task cost that previously would have required Opus.

Below the visible customer layer, Oracle, Broadcom and AMD play indirect roles. Broadcom designs custom AI silicon and is part of Anthropic's expanded partnership set; AMD's MI-series accelerators remain an option for non-frontier training and inference workloads; Oracle is a known Anthropic cloud and infrastructure partner through the broader Stargate ecosystem. None of these names move directly on Sonnet 5.5 pricing, but each captures a share of the spend Anthropic redirects toward compute.

Days to quarters: the immediate catalysts are (1) Haiku 5.5, which Anthropic committed to shipping 'in the coming weeks' after Opus 5.5, and whose price point versus GPT-6 Luna will determine the budget-tier narrative for the S-1; (2) the Anthropic S-1 public filing in mid-October, where the disclosed revenue trajectory, customer concentration and lock-up structure will reset the $2T valuation debate; and (3) OpenAI's likely counter-move on the next-generation frontier, where GPT-6 Astra is already at $10/$50 and a successor could either re-widen the gap or pull the frontier down toward Opus 5.5's $4/$20.

One to three years: the structural argument is that Anthropic can keep cutting per-token prices faster than its inference costs fall, because it has locked in 14+ gigawatts of capacity across four cloud providers at multi-year commitments, with Trainium2/3/4 providing a cost-controlled ramp on the AWS side. The risk is that the $2 trillion valuation requires revenue to keep scaling on a per-token price that keeps deflating — a compounding equation that breaks if the per-token decline slows before volume compensates. A second risk is OpenAI's larger installed base and partnership footprint; even with Sonnet 5.5 matching GPT-6 Sol on price, OpenAI still benefits from first-mover ChatGPT distribution and from the embedded Microsoft footprint. A third risk is the circular-financing question around NVIDIA's three-role deal, which could surface in the S-1 risk factors and pressure the multiple independent of any model launch.