# Anthropic Ships Claude Sonnet 5.5 at Unchanged Prices and Claims 30 Percent Lower Cost Per Task

Third-party roundups put the model at 70.6 percent on Terminal-Bench 4.0, above the larger Opus 5.5 at 66.4 percent, while Opus still leads on SWE-Bench Pro.

- Published: 2026-09-29T06:12:03.689Z
- Canonical: https://polylog.news/ai/2026-09-29/anthropic-ships-claude-sonnet-5-5-at-unchanged-prices-and-cl
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic](https://www.anthropic.com/claude-sonnet-5-5), [Polylog editors](https://polylog.news)

Anthropic released [Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) on Sunday, six days after Opus 5.5. The company describes it as a model that runs 30 percent faster and costs up to 30 percent less for most work. Token prices are unchanged at $2 per million input tokens and $10 per million output tokens, so the savings come from using fewer tokens per task rather than from a price cut.

The positioning is explicit about scope. Russian-language coverage of the launch [summarised it](https://t.me/ai_machinelearning_big_data/11034) as the fast, inexpensive complement to Opus 5.5 for well-specified everyday work such as bug fixes, documents, slides and spreadsheets. Anthropic's own framing matches that description.

The benchmark numbers in circulation are worth separating by source. [One roundup](https://www.marktechpost.com/2026/09/28/anthropic-releases-claude-sonnet-5-5-70-6-on-terminal-bench-4-0-at-the-same-2-10-price/) reports Sonnet 5.5 scoring 70.6 percent on Terminal-Bench 4.0 against 66.4 percent for Opus 5.5, 81.3 percent on SWE-Bench Pro against 89.9 percent for Opus 5.5 and 63.2 percent for Sonnet 5, and 90.3 percent on SWE-Bench Multilingual. A cheaper model outscoring its larger sibling on an agentic terminal benchmark, while trailing it by more than eight points on repository-level coding, looks like a harness advantage rather than a broad capability gain. Neither figure has an independent reproduction yet.

## What this means

The mechanism is cost per completed task, not cost per token, and that is where the mid-tier model class now competes. Anthropic gains if teams shift routine agent traffic from Opus to Sonnet without changing their code, because switching within a family costs almost nothing while switching vendors costs evaluation time. Exposed are the pricing tiers at OpenAI and Google that were positioned against Sonnet 5 economics, and any inference startup whose pitch was cheaper service on a weaker model. The precedent is the Haiku and Flash tiers, whose arrival pulled most production traffic off the flagship models within a few quarters. The more likely outcome is that the next mid-tier release from OpenAI or Google gets marketed on tokens consumed per task rather than on list price, within weeks rather than months. Independent Terminal-Bench reruns that fail to reproduce the 70.6 percent figure would undercut this account.

## What to watch

- Whether independent evaluators reproduce the Terminal-Bench 4.0 gap between Sonnet 5.5 and Opus 5.5, which would show whether it reflects a harness effect or a genuine agentic gain.
- Whether OpenAI or Google respond with a mid-tier release priced on tokens per task, a sign that the cost metric has shifted industry-wide.
- Whether Anthropic publishes the token-consumption data behind its 30 percent cheaper claim, since the price per token itself did not change.