Anthropic's announcement for Claude Sonnet 5 includes a quiet admission: tasks that used to stall mid-workflow on Sonnet 4.6 now finish. "We handed Claude Sonnet 5 a two-part job - update Salesforce account tiers, send a launch announcement to enterprise contacts - and it finished end to end. That used to stall halfway." That is the clearest single sentence in the launch post, and it is the right frame for what this release actually is: not a better chat model with incremental benchmark scores, but a Sonnet-class model that has crossed a threshold for agentic follow-through.
Claude Sonnet 5 was released June 30, 2026, and Anthropic positions it as the most agentic Sonnet model so far - designed to bring more Opus-like coding, tool use, browser work, and professional-task performance into a cheaper Sonnet-class tier.
From that date, it became the default model for Free and Pro plans and launched on the Claude Platform with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which it moves to $3 in and $15 out.
Those numbers matter because the prior incumbent - Sonnet 4.6 - was priced at $3 in and $15 out. The introductory window makes Sonnet 5 materially cheaper right now, which is the migration window Anthropic is clearly trying to open. The catch, and it is a real one, is that this is not a drop-in swap.
What actually changed in Sonnet 5
Sonnet 5 can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models. That improvement is real, but the mechanism matters. Sonnet 5 improves the Sonnet line for coding, agentic work, tool use, computer use, and professional knowledge work, and also changes the operating surface through adaptive thinking, effort levels, a new tokenizer, request-compatibility constraints, and cyber-safety behavior.
The most immediately disruptive of those changes is the tokenizer. Anthropic's published notes say Sonnet 5 uses an updated tokenizer that can increase token counts by roughly 1.0-1.35x depending on content, and the company indicates the introductory pricing is intended to be roughly cost-neutral during the transition. "Roughly cost-neutral" is doing a lot of work in that sentence. This is offset by an average 30% increase in tokens for equivalent text, which Sonnet 5 owes to the new tokenizer; launch documents mention that with the discounted price, Sonnet 4.6 workloads should cost the same when run on Sonnet 5. Any workflow with per-request budget enforcement built around Sonnet 4.6 token counts will underestimate costs until recalibrated.
The second operational change: if your codebase sets temperature, top_p, or similar parameters, those calls will error on Sonnet 5 - one of the three breaking changes most likely to silently break agent loops in production. Adaptive thinking is also on by default, which affects latency and token spend in ways that are subtle if you have not explicitly accounted for effort levels.
Sonnet 5 vs Opus 4.8: where the gap still matters
Sonnet 5's performance is close to that of Opus 4.8, but at lower prices. Close, not equal. The gap is specific. Sonnet 5's agentic coding benchmark lands 5 points above Sonnet 4.6 and 6 points below Opus 4.8.
Sonnet 5 is a good upgrade for workloads running on Sonnet 4.6 that would benefit from stronger coding, tool use, and agentic follow-through; Opus 4.8 remains the higher-capability reference for complex reasoning, long-horizon agentic coding, and high-autonomy work, but the gap is significantly smaller.
That residual gap is not randomly distributed. Sonnet 5 shows an overall lower rate of undesirable behaviors than Sonnet 4.6 in Anthropic's safety assessments and is generally safer to use in agentic contexts; evaluations also show it has a much lower ability to perform cybersecurity tasks than current Opus models. For most product teams, the cybersecurity ceiling is irrelevant. For a small subset - penetration testing tooling, red-teaming agents, automated security audits - Opus 4.8 remains the only option.
The price differential between the two tiers is also significant. Anthropic lists Opus 4.8 at $5 per million input tokens and $25 per million output tokens. At Sonnet 5's introductory rate, you are paying 40% of the Opus input price and 40% of the output price. For a multi-step agent workflow that burns 2 million tokens per task, the difference between Opus and Sonnet compounding across thousands of daily runs is the kind of bill that becomes a board conversation. For engineers designing production agents, that means more choices where mid-tier models can deliver acceptable autonomy at materially lower inference cost than Opus-tier or frontier models.
The honest position is this: Sonnet 5 is the right default to test first for most agentic workloads. Opus 4.8 is where you go when Sonnet 5 demonstrably breaks down on your task.
What to actually validate before migrating
The case for moving is straightforward if you are on Sonnet 4.6 and your agents are hitting the ceiling on multi-step tasks. Feedback from Anthropic's early access partners was consistent: Sonnet 5 is much more agentic than its predecessors - testers described how it finishes complex tasks where previous Sonnet models would stop short, how it checks its own output without being asked, and how it handles sustained coding, tool use, and debugging well across messy technical contexts.
But the migration is not a model ID change. It is a validation exercise with three specific checkpoints.
First, run end-to-end agent loop tests before switching production traffic - not single-turn accuracy checks, but full workflow runs that exercise your actual tool chains. Independent analysis found that agentic tasks can still cost more per task despite lower headline token prices; better plan-keeping and fewer correction loops matter for practitioners because they reduce orchestration overhead and end-to-end latency in agent pipelines. The token savings only materialize if the model actually completes tasks in fewer turns.
Second, audit every place in your codebase where you pass temperature, top_p, or other sampling parameters. Those calls will now throw errors.
The migration checklist: update your model ID to claude-sonnet-5; audit and remove temperature and sampling parameters; recalibrate token budget enforcement against Sonnet 5's tokenizer; and run agent loop tests end to end before switching production traffic.
Third, re-examine any cost model built on Sonnet 4.6 token counts. The tokenizer change is not catastrophic, but a 1.35x multiplier on text-heavy inputs can move a carefully calibrated budget by thousands of dollars a month at scale.
The broader signal
The arrival of Claude Sonnet 5 continues an industry trend where agentic capability becomes a baseline expectation, and the primary commercial differentiation shifts toward cost, reliability, and integration. That is the correct read. A Sonnet-tier model that can genuinely run autonomous multi-step workflows changes the unit economics of agents for any team paying Opus prices today.
The practical question for most teams is not whether Sonnet 5 is impressive in isolation, but when to use it versus Opus 4.8. Anthropic's own summary: "Opus 4.8 is still the model of choice for higher accuracy on these tasks, but Sonnet 5 provides developers with lower-priced options that are of much higher quality than what was previously available. Between Sonnet 5 and Opus 4.8, users can adjust the effort level to find the right balance of cost and performance."
An AI teammate like Beagle - running inside Slack or Teams, summarizing threads, drafting standups, surfacing context - sits squarely in the Sonnet 5 sweet spot: multi-step, tool-using, professional knowledge work that benefits from better follow-through without needing the full Opus ceiling. The model that handles those workflows improved materially this week.
The evaluation you need to run is not "is Sonnet 5 better?" It almost certainly is. The question is whether it is better enough for your specific failure cases to justify the migration work - and whether eight weeks is enough time to find out before the price goes up. For most teams running agents today, the answer to both is yes.