Anthropic launched Claude Sonnet 5 on June 30, making the new model available across Claude plans, Claude Code, and the Claude Platform with pricing designed to push more agentic work into day-to-day production.
The company is positioning Sonnet 5 as a cheaper execution layer for coding agents, browser-using agents, research workflows, and professional automation rather than as its most powerful model overall. Anthropic says the model closes much of the gap with Opus 4.8 on agentic tasks while costing less to run, especially during an introductory pricing window that lasts through August 31, 2026.
For readers who build with Claude, the launch matters less as a benchmark headline than as an operating-cost and migration decision. Sonnet 5 is now the default model for Claude Free and Pro users, is available to Max, Team, and Enterprise customers, and can be called through the API as claude-sonnet-5, according to Anthropic’s announcement. GitHub also made Claude Sonnet 5 generally available in GitHub Copilot on launch day, giving software teams another route to test it inside existing coding workflows.
The Pricing Is the Point
Sonnet 5 starts at $2 per million input tokens and $10 per million output tokens through August 31. After that, standard pricing moves to $3 per million input tokens and $15 per million output tokens. Anthropic lists Opus 4.8 at $5 per million input tokens and $25 per million output tokens, which gives teams a clear reason to test whether Sonnet 5 can carry workflows that previously needed the larger model.
That price gap is especially relevant for agents. A normal chatbot exchange may be short, but an agent that searches, reads files, calls tools, writes code, checks results, and retries can burn through large context and output budgets quickly. If Sonnet 5 can finish more of those jobs without escalating to Opus, teams get a more plausible cost structure for letting agents run repeatedly across codebases, support queues, data analysis tasks, and internal operations.
There is a catch in the accounting: Anthropic says Sonnet 5 uses an updated tokenizer, and the same input can produce roughly 1.0 to 1.35 times as many tokens depending on the content. The launch discount is meant to soften that transition, but developers should still measure real prompts before assuming a clean bill reduction. Long repository context, generated logs, verbose tool traces, and large documents may not all move at the same effective price.
What Developers Need to Check
The Claude Platform release notes add several details that matter for migration. Sonnet 5 supports a 1 million token context window and 128,000 maximum output tokens, and it keeps the same broad tool and platform feature set as Sonnet 4.6, except Priority Tier is not available on Sonnet 5.
API behavior also changes. Adaptive thinking is on by default. Manual extended thinking in the older thinking: {type: "enabled", budget_tokens: N} form has been removed and returns a 400 error. Requests that set temperature, top_p, or top_k to non-default values also return a 400 error. Those are small details in a launch post, but they are the kind of details that can break wrappers, evaluation harnesses, and production routing logic if teams swap model names without testing.
The practical migration checklist is straightforward: run token counts on representative prompts, test refusal and error handling, check any custom sampling settings, rerun coding-agent evaluations against real pull requests, and compare medium-effort Sonnet 5 against Opus 4.8 before changing default routes. Teams using Claude Code or Copilot should also watch whether the model’s stronger follow-through changes review load. A model that completes more work can still create more diffs to inspect.
Anthropic Is Selling Follow-Through
Anthropic’s examples emphasize agentic behavior: sustained coding, browser and terminal use, debugging, legal research, customer operations, insurance workflows, and live data exploration. The common thread is not a single new modality. It is follow-through across messy tasks where a model needs to plan, use tools, check output, and recover when the first attempt fails.
That is also where the model market is becoming more practical. Many companies have already tested agents in demos. The expensive part is letting them run often enough, with enough context and tool use, to become part of ordinary work. Sonnet 5 is Anthropic’s answer to that deployment problem: keep Opus 4.8 for harder jobs, but make the mid-tier model capable enough that agents do not constantly hit a quality ceiling.
GitHub Copilot availability makes that test easier for software teams, because developers can compare Sonnet 5 against other models in a familiar environment rather than building a separate evaluation setup first. The same logic applies to Claude Code and the API: the important measurement is not one impressive prompt, but whether the model consistently finishes work that saves human time after review.
The Safety Framing Is Narrower Than the Capability Story
Anthropic says its safety assessments found Sonnet 5 generally safer than Sonnet 4.6 in agentic contexts, including better refusal of malicious requests and stronger resistance to prompt-injection attempts. The company also says the model has lower hallucination and sycophancy rates than Sonnet 4.6.
Cybersecurity is treated carefully. Anthropic says it did not deliberately train Sonnet 5 on cyber tasks, and that the model performs substantially below Opus 4.8 and Mythos 5 on dangerous cyber capability tests. In one Firefox exploit evaluation cited by the company, Sonnet 5 did not produce a working exploit, though it had slightly more partial success than Sonnet 4.6. Anthropic launched it with cyber safeguards enabled by default and includes it in the Cyber Verification Program for organizations already enrolled.
That distinction matters for enterprise buyers. Sonnet 5 is meant to be broadly useful for agents, coding, and professional work, not a replacement for restricted high-capability cyber models. Security teams should still treat it as an agentic system that can touch tools, code, browsers, and internal data. The risk profile depends on the permissions around the model as much as the model itself.
What to Watch Next
The first useful signal will come from real migration data. If teams can move a meaningful share of coding-agent and workflow-agent traffic from larger models to Sonnet 5 without losing reliability, Anthropic will have a stronger case that agents can become routine infrastructure instead of expensive demos.
The second signal is cost after the launch window. The August 31 pricing deadline gives developers two months to benchmark the model under real workloads. After that, the standard rate still sits below Opus 4.8, but the tokenizer change and higher token usage in long-running agents will make actual bills uneven across use cases.
The third signal is control. As agent models get cheaper and more capable, organizations will need clearer routing rules, permission boundaries, audit logs, prompt-injection defenses, and human review paths. Sonnet 5 may make agents cheaper to run, but it also makes the surrounding governance harder to ignore.
