ModelsArchitectures & capability
Anthropic launches Claude Sonnet 5.5 at $2/$10 per million tokens, but max-effort runs cost more per task than Opus 5.5 in early tests
Anthropic's Claude Sonnet 5.5 keeps Sonnet 5 pricing and claims faster, cheaper tasks. Early Artificial Analysis tests found that its maximum effort setting used so many tokens that it cost more per task than Opus 5.5.
QuietHow it’s measured
YouTube video · 81 views · 1 like
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisSonnet 5.5 launched on Sep 28, 2026 at the same $2 input and $10 output per million tokens as Sonnet 5. Anthropic reports that output is over 30% faster and that tasks cost up to 30% less. It credits the savings to fewer tokens and tool calls and to recalibrated effort levels, not to a price cut. [6] [7] [4]
At max effort, Artificial Analysis measured about 193,000 output tokens per task. That works out to roughly $7.60 per task, against about $5.98 for Opus 5.5 and about 50% more than Sonnet 5. The test ran on a pre-release build with a bug that has since been fixed, and a re-test is planned. [4] [8]
The measurements show that cost per task depends heavily on the effort setting, so per-token price alone is misleading. Before switching from Opus 5.5 or promising savings, teams should benchmark Sonnet 5.5 at the effort level they will actually use.
Executive brief
Anthropic's Claude Sonnet 5.5 is cheaper per token than Opus 5.5. At maximum effort, though, Artificial Analysis found it cost more per task: about $7.60 against $5.98 for Opus 5.5, because it wrote about 193,000 tokens per task (DEV Community). The commentary under review, a Reddit post and a YouTube video presented by an AI avatar of Julian Goldie, repeats Anthropic's headline claims.
Read the full section
Anthropic's Claude Sonnet 5.5 is cheaper per token than Opus 5.5. At maximum effort, though, Artificial Analysis found it cost more per task: about $7.60 against $5.98 for Opus 5.5, because it wrote about 193,000 tokens per task (DEV Community). That run used a pre-release build with a bug that has since been fixed, and Artificial Analysis plans to re-test. The commentary under review, a Reddit post and a YouTube video presented by an AI avatar of Julian Goldie, repeats Anthropic's headline claims. These are "30%+ faster" and "up to 30% less" per task, plus a temporary 50% usage discount in the Claude app. The commentary does flag the footnotes on these claims, but it also pushes a paid community.
What changed and event timeline
Opus 5.5 opens the 5.5 family
Anthropic released
claude-opus-5-5at $4/$20 per million tokens, according to aggregator reports. Primary documentation for this date was not reviewed ().Sonnet 5.5 launches
Anthropic claims output is 30%+ faster than Sonnet 5 and tasks cost up to 30% less, at the same $2/$10 pricing (;).
Artificial Analysis finds heavy token use
At max effort, cost per task was about 50% higher than Sonnet 5. The results came from a pre-release build with a structured-outputs bug, and re-runs are planned ().
App usage promotion starts
After Claude creates or edits a design, deck or doc, the follow-up work in that chat counts 50% less toward usage limits until Oct 15, 11:59 PM PT (;).
Capabilities and access
- Model:
claude-sonnet-5-5. It is available in the Claude apps, the API, AWS, Google Cloud and Azure (Anthropic). - Pricing: $2 per million input tokens and $10 per million output tokens. Cache reads cost $0.20 per million (SiliconANGLE).
- Effort levels: low, medium (the default in the Claude apps), high (the default on the platform), xhigh and max (Anthropic).
Read the full section
- Model:
claude-sonnet-5-5. It is available in the Claude apps, the API, AWS, Google Cloud and Azure (Anthropic). - Pricing: $2 per million input tokens and $10 per million output tokens. Cache reads cost $0.20 per million (SiliconANGLE).
- Effort levels: low, medium (the default in the Claude apps), high (the default on the platform), xhigh and max (Anthropic).
- Positioning: Anthropic pitches it for well-scoped everyday tasks, bug fixes, documents, slides and spreadsheets. Opus 5.5 is pitched for complex work that needs careful judgment (9to5Mac).
Technical analysis for researchers and developers
- Where the savings come from: Anthropic credits fewer tokens and fewer tool calls, not a price cut.
- Effort is the main cost variable: at max effort, the model used about 193k output tokens per task.
- Reproducibility: all of the Artificial Analysis evaluations ran on the buggy pre-release deployment, so treat the decimals as provisional until the re-runs are out.
Read the full section
- Where the savings come from: Anthropic credits fewer tokens and fewer tool calls, not a price cut. It also says effort levels were recalibrated, so a given level no longer means the same amount of thinking it did on Sonnet 5 (DEV Community).
- Effort is the main cost variable: at max effort, the model used about 193k output tokens per task. That is roughly 60% more than Opus 5.5 or Sonnet 5 at max effort (Adyog Pulse).
- Reproducibility: all of the Artificial Analysis evaluations ran on the buggy pre-release deployment, so treat the decimals as provisional until the re-runs are out.
- Implication for developers: compare cost per task at the effort level you will actually use. Per-token price alone will mislead.
Claims and evidence
- Over 30% faster than Sonnet 5
- Faster than Opus 5.5
- Up to 30% less per task
Read the full section
| Claim | Evidence |
| Over 30% faster than Sonnet 5 | Anthropic-reported; no like-for-like independent test found (Anthropic). |
| Faster than Opus 5.5 | The video cites about 137 vs 93 tokens/s (01:06). The Artificial Analysis comparison page currently shows 139 vs 97 (Artificial Analysis). The figures change with effort level and over time. |
| Up to 30% less per task | Anthropic-reported. It conflicts with the higher max-effort cost Artificial Analysis measured. |
| Box tasks 2.4x faster | Box says Sonnet 5.5 was also more accurate and used 12% fewer tokens. This is a customer quote on Anthropic's own page (Anthropic). |
| Terminal-Bench 4.0 | Anthropic reports 70.6% (The Decoder). Artificial Analysis measured 63.6%, which is still above Opus 5.5's 59.6% (DEV Community). |
Context and prior work
- A step up from Sonnet 5: Anthropic reports Terminal-Bench 4.0 rising from 10.3% to 70.6% and Chartography from 15.6% to 61.6% (The Decoder).
- Close to Opus 5.5 overall: on the Artificial Analysis Intelligence Index, Sonnet 5.5 scores 56 against Opus 5.5's 58 (Artificial Analysis).
- Smaller model still to come: Haiku 5.5 is expected within weeks (SiliconANGLE).
Read the full section
- A step up from Sonnet 5: Anthropic reports Terminal-Bench 4.0 rising from 10.3% to 70.6% and Chartography from 15.6% to 61.6% (The Decoder).
- Close to Opus 5.5 overall: on the Artificial Analysis Intelligence Index, Sonnet 5.5 scores 56 against Opus 5.5's 58 (Artificial Analysis).
- Smaller model still to come: Haiku 5.5 is expected within weeks (SiliconANGLE).
Limitations, safety and contested findings
- Factual accuracy: on AA-Omniscience, Sonnet 5.5 scored 54% against Opus 5.5's 66% (Adyog Pulse).
- Safety measures: this is the first Sonnet with cyber safeguards and fallback behaviour for higher-risk requests.
- Promotion limits: the discount covers only the 5-hour session window, not weekly limits. Reports disagree on whether Cowork is covered.
Read the full section
- Factual accuracy: on AA-Omniscience, Sonnet 5.5 scored 54% against Opus 5.5's 66% (Adyog Pulse).
- Safety measures: this is the first Sonnet with cyber safeguards and fallback behaviour for higher-risk requests. Anthropic also added classifiers against distillation and invisible text watermarking (SiliconANGLE).
- Promotion limits: the discount covers only the 5-hour session window, not weekly limits. It excludes the Free and Enterprise plans, Claude Code and the API (SmartScope). Reports disagree on whether Cowork is covered.
- The commentary is promotional: the video is presented by a digital avatar (00:33) and includes repeated pitches for a paid community (01:39).
Business and practitioner implications
Read the full section
- Default to Sonnet 5.5 at low to high effort for routine work. Benchmark the max setting yourself before using it, because that is where its cost per task overtook Opus 5.5.
- Use the promotion before it ends on Oct 15. Keep edits in the same chat, as the video advises (02:48). The discounted rate lasts for 10 messages after each artifact edit.
- Plan for review time. Anthropic says slides "need minimal editing", and the video notes that minimal does not mean zero (03:37).
- Wait for the re-test. Don't quote the 30% savings figure to clients until Artificial Analysis publishes its re-run.
The source trail.
Sources (13)
Is the Claude Sonnet New Model Worth It? The Honest Numbers
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=NkzDcwZh0N8); automatic text may contain errors.
reddit.comClaude Sonnet 5.5 Is 30% Faster, and That's Not the Best Part
Related coverage; assess separately
reddit.comLinked video: Is the Claude Sonnet New Model Worth It? The Honest Numbers
linked video auto captions
www.youtube.com