Anthropic has turned what was meant to be a temporary launch discount into the permanent price of Claude Sonnet 5. The model now costs $2 per million input tokens and $10 per million output tokens, according to an August 10 update to Anthropic’s original announcement.
That change matters because Sonnet 5 is positioned as the company’s practical model for coding, tool use and autonomous workflows. For teams deciding whether an AI agent can move from a prototype into production, a predictable lower price can be as important as a higher benchmark score.
Background: where Claude Sonnet 5 fits
Anthropic launched Claude Sonnet 5 on June 30, 2026 as the successor to Sonnet 4.6. It sits below the company’s most capable and more expensive models, but targets the high-volume work where speed, capability and cost must be balanced.
The company describes Sonnet 5 as its most agentic Sonnet model so far. In practical terms, that means it is designed to plan multi-step jobs, call tools, browse, work in terminals, write and debug code, and continue a workflow with less hand-holding. Anthropic says its performance is close to Opus 4.8 on some tasks, although results depend on the task and the selected reasoning effort.
Sonnet 5 is available in Claude’s consumer plans and through the Claude API. Anthropic also lists availability through Amazon Web Services, Google Cloud and Microsoft Foundry, giving enterprise buyers several deployment routes.
What changed in Claude Sonnet 5 pricing?
At launch, Anthropic announced an introductory API price of $2 per million input tokens and $10 per million output tokens. A higher standard rate of $3 for input and $15 for output had been scheduled to begin on September 1.
Anthropic’s August 10 changelog now says the introductory rate is permanent and that the planned $3/$15 pricing no longer applies. This is effectively a one-third reduction from the previously planned standard rate for both input and output.
A simple cost example
Consider a monthly workload that sends 100 million input tokens and generates 20 million output tokens. At the permanent $2/$10 rates, the base API cost would be $400: $200 for input and $200 for output. At the abandoned $3/$15 rates, the same usage would have cost $600.
Real bills can differ because of prompt caching, batch processing, cloud-provider charges and other API features. Anthropic says prompt caching can reduce relevant costs by up to 90%, while batch processing can offer a 50% saving. Teams should model their own input-to-output ratio rather than relying on one headline figure.
Why the permanent price matters
AI agents often consume far more tokens than a single chatbot answer. An agent may inspect files, read logs, call tools, review its own work and retry failed steps. Each pass can add context and output, so token prices multiply quickly across thousands of tasks.
Making the lower rate permanent removes a looming price increase from budgets. It also strengthens Sonnet 5’s appeal for businesses comparing models for customer support, software development, research, document processing and browser-based automation.
The move may also increase competitive pressure across the AI market. Model buyers increasingly compare total task cost rather than price per token alone. A cheaper model is not economical if it needs repeated retries, but a capable model that completes work in fewer steps can lower the full cost of a workflow.
Practical impact for developers and businesses
Recalculate production budgets
Teams that budgeted for the September increase can now revise forecasts. Use real telemetry: average input tokens, generated tokens, cache hit rates, retries, tool calls and successful task completion. These figures give a more useful cost estimate than a simple request count.
Test effort settings
Anthropic lets API users control the model’s thinking effort. Higher effort can improve results on difficult work but may consume more tokens and take longer. A sensible rollout tests low or medium effort for routine tasks, then escalates only when complexity or risk justifies it.
Keep model selection flexible
Sonnet 5 may be a strong default for coding agents and general automation, but it should not automatically replace every model. Lightweight classification, extraction or routing may be cheaper on a smaller model. The hardest reasoning tasks may still justify a more capable tier. Routing requests by difficulty can produce better economics than forcing one model to handle everything.
Evaluate migration before switching
Developers can call the model using the claude-sonnet-5 identifier. Before changing production traffic, test instruction following, tool schemas, structured outputs, latency, refusal behaviour and regression cases. Agentic systems can fail in ways that ordinary chat tests do not reveal, particularly when they can modify files or act on external services.
Risks, limitations and concerns
Anthropic’s performance and safety statements are based partly on its own evaluations. Independent testing in a team’s real environment remains essential. Benchmark gains do not guarantee reliable performance on a particular codebase, browser workflow or business process.
Autonomous tool use also expands the impact of mistakes. Businesses should apply least-privilege access, isolate sensitive systems, log actions, require approval for high-impact steps and defend against prompt injection from web pages, emails and documents.
Cost can still become unpredictable when agents enter loops, ingest very large contexts or repeatedly retry tools. Hard token limits, timeouts, spending alerts and per-task budgets should be built into production deployments. Buyers using a cloud marketplace should also confirm the exact price and regional terms shown by their provider.
What to watch next
The next important signal will be whether the permanent price encourages broader use of long-running agents and whether competitors respond with price or capability changes. Developers should also watch Anthropic’s model documentation and cloud-provider listings for any differences in availability, rate limits or billing.
More importantly, teams should track cost per successful outcome. The winning model for an agent is not necessarily the one with the lowest token rate or the best single benchmark; it is the one that completes a defined job reliably, safely and at an acceptable total cost.
Conclusion
Anthropic’s decision to make Claude Sonnet 5’s $2 input and $10 output rates permanent is a meaningful change for anyone building token-intensive AI workflows. It eliminates a planned 50% price increase from the current rate and gives developers a more stable basis for production planning.
The lower price improves the model’s value proposition, but disciplined testing still matters. Measure completed-task cost, tune effort levels, limit agent permissions and retain human approval where errors could cause financial, security or reputational harm.