Claude Sonnet 5.5: 30% Faster, 30% Cheaper for Enterprise
Claude Sonnet 5.5: 30% Faster, 30% Cheaper for Enterprise
Anthropic released Claude Sonnet 5.5 on September 28, 2026, the second model in the Claude 5.5 family following Claude Opus 5.5 from the previous week. The short version for enterprise teams: it runs 30% faster than Claude Sonnet 5, costs up to 30% less per task at the same per-token rate, and scores 70.6% on Terminal-Bench 4.0 agentic coding, up from Sonnet 5’s 10.3%.
Sonnet 5.5 is available today across all major cloud platforms. For teams already running Sonnet 5 in production, the migration path involves five breaking API changes detailed below.
What Changed and What the Numbers Mean
Anthropic’s positioning is clear: Sonnet 5.5 is the everyday workhorse of the Claude 5.5 family. Opus 5.5 handles work that requires judgment at every step. Sonnet 5.5 handles work where the approach is already defined and what remains is to execute it quickly and cheaply.
The Terminal-Bench 4.0 jump, from 10.3% to 70.6%, is the most striking benchmark result. Terminal-Bench measures agentic coding performance on open-ended tasks. A 60-point gain in one generation is not incremental. Anthropic’s own system card describes the gap as a “significant improvement” while noting the model does not reach Opus 5.5 capability on the hardest reasoning tasks.
The model is also the first Sonnet to complete Pokémon Red working entirely from screenshots, a proxy for sustained visual reasoning across many steps. This matters for enterprise workflows involving document analysis, UI testing, and any task where the agent must track visual state over a long horizon.
| Benchmark | Sonnet 5 | Sonnet 5.5 | Notes |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 10.3% | 70.6% | Source: Anthropic system card |
| GDPval-AA (real-world work quality) | Below Opus 5.5 | 2 pts below Opus 5.5 | Cross-occupation evaluation |
| Cyber challenge completion | 0.7% | 46.1% | Offensive eval; safeguards in place |
| Pokémon Red (vision/long-horizon) | Did not complete | Completed | First Sonnet to pass |
Atlassian reported 30% faster agent runs in their Rovo Agents product after early testing. Anthropic says enterprise customers account for approximately 80% of its business, with Salesforce, Databricks, Goldman Sachs, and Novo Nordisk among active users of the 5.5 family.
Pricing: Same Per Token, Fewer Tokens
Sonnet 5.5 carries the same published rates as Sonnet 5. The cost reduction comes entirely from the model consuming fewer tokens per task.
| Charge | Price |
|---|---|
| Input | $2.00 per million tokens |
| Output | $10.00 per million tokens |
| Cache write (5 minutes) | $2.50 per million tokens |
| Cache write (1 hour) | $4.00 per million tokens |
| Cache read | $0.20 per million tokens |
| Batch API | 50% discount on input and output |
US-only inference (inference_geo: "us") adds a 1.1x multiplier. Batch API pricing is unchanged.
The per-task cost reduction is not uniform. Teams running continuous, scoped agentic tasks at high volume, alert triage, SQL generation, document review, and short coding loops, will see the largest gains. Teams running occasional, open-ended queries will see smaller savings because the efficiency improvement compounds over many steps.
The cache read price at $0.20 per million tokens matches Opus 5.5’s rate and is 60% lower than Opus 5’s $0.50. For agentic workflows where a large cached system prompt dominates costs, this matters regardless of which model you use.
Five Breaking API Changes Before You Migrate
Five code changes break existing Sonnet 5 integrations. All were also required for Opus 5.5.
1. Disable up-front thinking with between_tools, not the old disable path. Setting thinking: {type: "disabled"} returns an error. The new setting between_tools turns off thinking before the first tool call while keeping the model’s reasoning in tool-calling loops. It works at high effort or below.
2. Remove forced tool use. tool_choice: {type: "any"} returns an error. If your integration requires a tool call on every turn, restructure the logic.
3. Thinking blocks are tied to the model and the conversation. Moving thinking blocks across accounts, including switching accounts mid-session in Claude Code, triggers the preserved thinking safeguard. This is the distillation protection Anthropic added to prevent industrial-scale capability extraction.
4. The computer_20251124 computer use tool returns a 400 error on the Claude API and Google Cloud. Replace it with {"type": "computer_toolset_20260801"}. Amazon Bedrock still accepts the older tool, but migration is recommended for consistency.
5. The advisor tool rejects earlier models. Claude Opus 4.8, Opus 4.7, and Sonnet 5 are no longer accepted as advisors. Update advisor model references before switching.
One additional change does not fail requests but affects streamed output: text between tool calls now arrives in thinking blocks with empty text at the default display setting, which makes streamed progress appear silent between calls. Set display: "minimal" or use between_tools to restore the previous behavior. Applications that display streaming output to users should test this before migration.
The full migration guide is in Anthropic’s documentation.
Cybersecurity Safeguards: First for a Sonnet Model
Sonnet 5.5’s cyber capability gain is substantial enough that Anthropic deployed the same blocking classifiers used for Opus 5.5 and Opus 5. High-risk cybersecurity requests fall back to Sonnet 5. Routine software development tasks, finding and fixing bugs, vulnerability scanning, security code review, are unaffected.
The 46.1% completion rate on offensive cyber challenges is a step change from Sonnet 5’s 0.7%. Anthropic describes this as “a significant improvement” while noting it still falls below Opus 5.5’s 67.6% and Fable 5.1’s 61.7%. Enterprise security teams with verified use cases can apply to Anthropic’s Cyber Verification Program for tiered access.
The model also adds distillation safeguards for the first time at the Sonnet tier. Reasoning extraction attacks that use large numbers of fake accounts to copy model capabilities at scale are now blocked.
The Full Claude Lineup After Sonnet 5.5
With Sonnet 5.5 in place, the current model tier covers most enterprise use cases. A third model, Claude Haiku 5.5, is scheduled “in the coming weeks” for high-volume, cost-sensitive applications.
| Model | Input / Output (per MTok) | Latency | Best for |
|---|---|---|---|
| Claude Fable 5.1 | $10 / $50 | Slower | Hardest problems, maximum intelligence |
| Claude Opus 5.5 | $4 / $20 | Moderate | Complex judgment, long-horizon agentic work |
| Claude Sonnet 5.5 | $2 / $10 | Fast | Well-scoped everyday tasks at scale |
| Claude Haiku 4.5 | $1 / $5 | Fastest | Current high-volume tier |
| Claude Haiku 5.5 | TBD | TBD | High-volume with 5.5 quality, coming soon |
The practical guidance from Anthropic’s own documentation: “For the hardest long-horizon work, an Opus model is the better choice.” Sonnet 5.5 fits when the task has a clear spec and a way to verify the result. Most enterprise automation pipelines that run continuously fall into this category.
What to Do Now
For teams on Sonnet 5, the five breaking changes above are the only blockers. Once addressed, Sonnet 5.5 is a drop-in upgrade with a better cost profile and meaningfully higher agentic coding performance.
For teams not yet on the 5.5 family, the Sonnet-to-Opus choice is now a straightforward scope question: scoped and repeatable tasks go to Sonnet 5.5, open-ended and judgment-intensive work goes to Opus 5.5. The pricing comparison with OpenAI’s GPT-6 models is worth reviewing for teams evaluating providers.
Model ID across all platforms: claude-sonnet-5-5. On Amazon Bedrock: anthropic.claude-sonnet-5-5. Available now on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
If you are building agentic systems on the 5.5 family and want to understand the right model split for your workloads, the Enera team is available to discuss.