Anthropic released Claude Sonnet 5.5 on September 28, the second model in its Claude 5.5 series, announcing the debut through an official post on X. The launch comes as the company builds toward an initial public offering, according to Reuters. It also underscores a change in the criteria used to evaluate AI technology: raw benchmark performance remains a key consideration, but companies increasingly weigh how expensive it is to deliver results with a specific model.

Anthropic says Sonnet 5.5 operates more than 30% faster than Sonnet 5 and can complete most tasks for up to 30% less, even though token prices remain unchanged. According to the model's launch page, the cost reduction comes from using fewer tokens to perform the same work.

Same sticker price, fewer tokens

Sonnet 5.5 is priced at $2 per million input tokens, $10 for output tokens, and $0.20 for cache-read tokens — the discounted rate for reusing context the model has already processed. Opus 5.5, which specializes in complex reasoning and judgment operations, costs $4 and $20 for input and output tokens, respectively. Sonnet is positioned for everyday workloads such as coding, document generation, and spreadsheets.

The pricing puts efficiency at the forefront of the discussion. For many tasks, businesses do not need the highest-performing model available; a low-cost model that consistently completes a job can matter more than scoring a few points higher on a benchmark. That dynamic is especially important for Anthropic, which, per Reuters, draws roughly 80% of its revenue from enterprise customers.

Benchmarks stop being the whole story

Anthropic says Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, an agentic benchmark that tests models on real terminal tasks, compared with 10.3% for Sonnet 5. But the Artificial Analysis leaderboard from independent evaluator Artificial Analysis shows why benchmark results are only part of the picture. At maximum effort, Sonnet 5.5 reaches a score of 56 on the intelligence scale, at an estimated cost of $7.60 per task. GPT-6 Sol achieves 48 points at $1.06 per task, while Xiaomi's MiMo-V2.6-Pro scores 46 for only $0.13. In other words, a model that leads on intelligence does not necessarily lead on cost efficiency — and for high-volume workloads, per-task gaps of that size compound quickly.

Those cost differences push companies toward multi-model strategies: harder tasks are assigned to premium-grade systems, while routine work is routed to cheaper alternatives. The approach lets organizations pay for top-tier capability only where it is genuinely needed.

The price of thought keeps falling

The shift fits into a much larger trend. According to an Epoch AI report published on September 22, the cost of achieving a given level of AI performance has fallen roughly 47% per quarter since 2023 — an approximately 13-fold yearly reduction. A separate Cryptopolitan report from September 29 states that corporate buyers increasingly favor the most affordable model that can complete the required work, while low prices also make it easier for businesses to justify running multiple models at once.

Cheaper tokens, bigger bills

Lower unit costs, however, do not ensure lower overall AI expenses. A Gartner report published on August 17, titled "Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028," describes an "inference paradox": falling token prices stimulate more complex workflows, which drives total token usage upward. Unit prices and total spending are different levers — the first sets the cost of each token, while the second depends on how many tokens a workflow consumes.

"Product leaders cannot rely on more efficient token economics to rationalize AI costs," said Will Sommer, Senior Director Analyst at Gartner, in the August 17, 2026 report.

Who actually gets the cheaper AI

According to Microsoft's Global AI Diffusion Report, low-cost and open-weight models have the potential to improve access to AI significantly, particularly in the Global South. Even so, lower model costs alone will not determine adoption — infrastructure, connectivity, and skills remain the deciding factors in whether cheaper AI is actually used.

For enterprise buyers, Sonnet 5.5 signals where the market is heading: the best model is not necessarily the most powerful one, but the one that can do the job well at a reasonable cost. Haiku 5.5, the next model in the series, due in the coming weeks, could push that shift even further by making high-volume AI work cheaper to perform — and by making it harder for businesses to justify paying premium prices for routine tasks.