NextFin News - Anthropic released Claude Sonnet 5.5 on Monday, a mid-tier model that runs more than 30% faster than its predecessor and completes most tasks for up to 30% less — at a sticker price the company is not raising. The launch lands six days after OpenAI cut its GPT-6 Sol and Luna API prices by half on a permanent basis, and it signals that the artificial-intelligence model race has pivoted from capability one-upmanship to the economics of every completed task.
The headline benchmark is startling for a mid-tier model: on Terminal-Bench 4.0, an agentic-coding evaluation, Sonnet 5.5 scores 70.6%, ahead of Anthropic's own flagship Opus 5.5 at 66.4% and far ahead of Sonnet 5 at 10.3%. Yet per-token pricing is unchanged from Sonnet 5 — $2 per million input tokens, $10 per million output tokens, $0.20 per million cache reads. Anthropic's argument is that the model needs fewer tokens to do the same work, so the effective cost per task falls even though the price per token does not.
This is the second launch in the Claude 5.5 family in six days. On September 22, Anthropic introduced Opus 5.5 at $4 per million input tokens and $20 per million output — 20% below Opus 5, the first price cut in the Opus line's history — and roughly ninety minutes later OpenAI responded with 50% cuts across its GPT-6 Sol and Luna tiers. By Monday, Google's Gemini 3.8 Flash was being offered at an introductory $0.75 per million input tokens and $3.75 per million output through the end of 2026. The sequence is the story: capability is no longer scarce enough to hold price.
The Mid-Tier Now Eats the Flagship
The most important number in Anthropic's release is not the 30% speed gain. It is the 70.6%.
Sonnet 5.5 is positioned explicitly as the complement to Opus 5.5, not its replacement. Anthropic frames Opus 5.5 as the model for "complex work requiring careful judgment," while Sonnet 5.5 is "strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets." That division of labor only holds if the flagship remains meaningfully stronger. On knowledge-work benchmarks, it still is — but by a margin now measured in single digits.
On GDPval-AA v2.1, a test of real-world work across occupations, Sonnet 5.5 scores 1,844 in Elo terms, two points behind Opus 5.5 at 1,846. On AA-Briefcase v1.1, the gap is 1,811 to 1,822. On FrontierCode 1.1, Opus 5.5 leads at 54.4% versus Sonnet 5.5's 46.2% at maximum effort — though at the higher "Xhigh" reasoning setting Sonnet 5.5 reaches 52.1%. But on Terminal-Bench 4.0, the agentic-coding test that has become the industry's shorthand for whether a model can actually ship code, the mid-tier model flips the hierarchy: 70.6% for Sonnet 5.5 versus 66.4% for Opus 5.5.
"Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, an agentic coding evaluation, compared to Sonnet 5's 10.3%. It scores two points below Opus 5.5 on GDPval-AA, a test of real-world work across a variety of occupations."
The practical implication is a cannibalization risk Anthropic is accepting deliberately. A developer running agentic coding workloads on Opus 5.5 at $4 per million input tokens is now paying double the per-token rate of a model that, on Anthropic's own benchmark, scores lower on that exact task. The company's hedge is that Opus 5.5 remains "clearly stronger at complex, open-ended work requiring sustained judgment" — but the set of tasks that genuinely require sustained judgment is the set that shrinks fastest as mid-tier models improve.
There is also a safety consequence. Because Sonnet 5.5's cybersecurity capabilities are "comparable" to Opus 5's, Anthropic says it is the first Sonnet model to launch with the same cyber safeguards and fallbacks that apply to Fable and Opus. Biology safeguards remain the same as Sonnet 5's. The company's framing is narrow — both safeguard classes target a limited set of high-risk requests, and routine software development and most life-sciences work are unaffected — but the move acknowledges that Sonnet-tier models are now capable enough to warrant flagship-level guardrails.
Price Is the New Battleground
The sequencing of the past week is what turns a product launch into a regime shift. Three moves, six days, three different labs:
- September 22: Anthropic cuts Opus list prices for the first time in at least four model generations — $4 input and $20 output per million tokens, 20% below Opus 5, with cache reads at $0.20 per million tokens, 60% below Opus 5.
- September 22: OpenAI cuts GPT-6 Sol and Luna API prices by 50%, to $2 per million input tokens and $10 per million output for Sol, and $0.10 and $0.50 for Luna. A company spokesperson confirmed the new rates carry no expiration date.
- September 28: Anthropic ships Sonnet 5.5 at the same $2/$10 sticker price as Sonnet 5 — pricing the company made permanent on August 10 after abandoning a planned September 1 reversion to $3/$15 — while delivering 30%+ faster generation and up to 30% lower cost per task.
Google is undercutting both on headline price. Gemini 3.8 Flash is offered at an introductory $0.75 per million input tokens and $3.75 per million output through December 31, 2026, with standard pricing scheduled to rise to $1.50 and $7.50 on January 1, 2027. Below the frontier tier, open-weight challengers price lower still. The result is a cost curve that slopes down on every axis: per-token price, per-task cost, and latency.
None of these cuts are framed as promotions by the labs that made them permanent. OpenAI's spokesperson explicitly confirmed the rates carry no expiration. Anthropic made Sonnet 5's $2/$10 introductory pricing permanent on August 10. When price cuts are permanent rather than promotional, they are not a tactical discount — they are a statement about the marginal cost of intelligence.
"The Opus line had never moved."
That observation, from Tomasz Tunguz, founder of Theory Ventures, captures why this week matters. The flagship tier had been the price anchor for the entire market; once it moves, every tier below it must follow.
The driver is on the supply side. Opus 5.5 "requires less compute to serve than Opus 5," Anthropic said in its launch note, and its pricing reflects that. For Sonnet 5.5 the company emphasizes faster generation and fewer tokens per task — fewer tool calls, shorter reasoning chains, better caching. Each is an inference-efficiency gain, and each compounds on the customer's bill: fewer tokens at an unchanged per-token price is what delivers the up-to-30% lower cost per task, while faster generation is a separate benefit for latency-sensitive workloads.
Cost Per Task, Not Cost Per Token
The metric that matters has changed, and the labs know it. For the past two years, the industry priced and compared models per million tokens — a unit that made sense when every request was a single prompt and a single response. It makes far less sense in an agentic world, where one user request can spawn dozens of model calls, tool invocations, file reads, and self-correction loops.
Anthropic's entire Sonnet 5.5 pitch is built on this shift. The per-token sticker price is unchanged at $2/$10 — the same price OpenAI now charges for GPT-6 Sol, its direct competitor. What Anthropic is selling is not cheaper tokens but fewer of them. In the company's own testing, Sonnet 5.5 completes the same task for up to 30% less than Sonnet 5 because it needs fewer tokens to get there. On OSWorld 2.1, a computer-use benchmark, it scores 80.1% partial completion versus Sonnet 5's 57.0%, and it is the first Sonnet model to beat Pokémon Red working only from screenshots — a proxy for long-horizon, trial-and-error work where wasted steps are the entire cost.
This reframing is defensive as much as it is descriptive. If the market priced intelligence strictly per token, a 50% price cut would be a 50% revenue cut for the same workload. Pricing per completed task changes the arithmetic: a lab can pass part of its efficiency gain to customers as a lower cost per task while retaining margin, and the lower cost per task expands the set of work that is economically worth automating. That is the Jevons paradox applied to inference — cheaper intelligence does not shrink the market; it widens the range of tasks a business will hand to a model.
The winners of that transition are not obvious. Application-layer companies that embed models into workflows — coding assistants, document automation, customer-support agents — see their gross margins expand as their largest input cost falls. Enterprises running high-volume, repetitive inference see immediate savings. The exposed parties are the labs themselves if they cannot grow volume fast enough to offset falling unit prices, and the software vendors whose products were priced on the assumption that AI would remain expensive enough to be a moat.
The Counter-Thesis: A Race to the Bottom
The strongest case against the structural-deflation reading is that this is a margin-destroying race to the bottom, and that the labs are cutting price because they have no other lever. The bear argument runs: training costs are still rising — Anthropic's CEO has projected that training leading models could cost up to $100 billion between 2026 and 2029 — while inference prices fall. If revenue per token falls faster than efficiency improves, gross margins compress, free cash flow for the next training run shrinks, and the pace of frontier progress slows. In that world, today's price cuts are not a sign of health but of desperation to defend share against OpenAI and a cohort of well-funded Chinese and open-weight challengers.
There is evidence for pressure. Software stocks sold off sharply through 2026 on the fear that AI agents will displace enterprise software — a portfolio review of the selloff put the sector's technology-software gauge down roughly 23% for the year, with some SaaS names down 30% to 50% since January. That selloff is a bet that AI will take revenue away from incumbents. The same logic, applied to the labs, says the revenue they take will be low-margin revenue.
The answer to the bear case rests on two points. First, the price cuts are being delivered by efficiency, not just discounting. A model that uses 30% fewer tokens per task reduces the compute required for that task by roughly a third; the revenue cut and the cost cut move together rather than squeezing margin. Second, and more important, demand for inference is not fixed. The constraint on AI adoption over the past two years has rarely been desire; it has been cost and latency. Remove both, and the volume of automatable work can expand faster than the price falls.
The falsifying signal is specific: if per-token API prices at the three frontier labs stabilize or rise for two consecutive quarters while inference efficiency plateaus, the structural-deflation thesis is wrong. A more direct read would come from the infrastructure layer — if the largest cloud and data-center customers report flat or declining spend per token while their token volumes grow, the deflation is real and accelerating. Watch those two metrics, not the headlines.
What Comes Next
The immediate catalyst is Haiku 5.5. Anthropic says its smallest model, built for high-volume and cost-sensitive applications, will join the Claude 5.5 family in the coming weeks. If Sonnet 5.5 delivers flagship-adjacent agentic performance at mid-tier prices, a Haiku that inherits even a fraction of those efficiency gains would push the cost floor lower still — and put direct pressure on the sub-$1-per-million-input-token tier where Google, DeepSeek, and open-weight models currently compete.
By time horizon, the picture splits. In the short term, the release is a sentiment event for a software market already repricing on AI-disruption fears; Anthropic is a private company, so there is no direct equity move, but the pressure transmits through the same channel that drove the 2026 software selloff. Over the medium term, the beneficiaries are application-layer vendors and enterprises with high-volume inference workloads, whose unit economics improve as cost per completed task falls. Over the long term, the question is whether the labs can keep volume growth ahead of price declines — and whether "cost per task" becomes the industry's pricing standard, replacing the per-token model that has governed the sector since the API era began.
Base case: cost per task keeps falling at a double-digit annual pace as efficiency gains compound, and the mid-tier models continue to absorb workloads that previously required flagship models. Upside case for the labs: Jevons dynamics dominate, token volumes more than offset price cuts, and revenue grows even as unit prices fall. Downside case: price competition outpaces efficiency, gross margins compress, and the capital intensity of frontier training forces consolidation among all but the best-funded labs.
The central judgment: this is not a cyclical discount cycle that will revert. Permanent price changes, permanent efficiency gains, and a competitive field that includes well-funded open-weight challengers make inference-cost deflation structural — it will not reverse on its own. The market is no longer paying for how smart a model is per token; it is paying for how much work gets done per dollar. Sonnet 5.5 is the first mid-tier model built explicitly for that world, and the flagship tier is the one now on the defensive.
Explore more exclusive insights at nextfin.ai.
