【2026/6/30】Claude Sonnet 5リリース — 1Mコンテキスト標準化・temperature廃止・実質3割の値上げもAnthropic released Claude Sonnet 5 on June 30, 2026, standardizing 1M-token…

Anthropic released Claude Sonnet 5 on June 30, 2026, and the update carries several changes that directly affect how developers build on the company's API. The most consequential are the standardization of a one-million-token context window, the deprecation of the temperature parameter, and what amounts to a roughly 30 percent increase in effective pricing. Taken together, these shifts require teams to revisit both their integration code and their cost projections, which is why the release is being treated as more than a routine model refresh.

The most immediately visible change is the context window. By making a one-million-token context standard across Sonnet 5, Anthropic brings its mid-tier model in line with the largest windows already available elsewhere in the market. Sonnet has historically occupied the middle of Anthropic's lineup, positioned between the lighter, faster Haiku models and the more capable Opus tier, and is often chosen for production workloads that need a balance of quality and cost. A larger default context allows applications to feed in longer documents, extended conversation histories, or larger codebases without elaborate chunking and retrieval workarounds. It is worth noting, however, that a bigger window does not automatically translate into better results across an entire prompt, and models can still weight information unevenly over very long inputs.

The deprecation of the temperature parameter is a more structural change to the API contract. Temperature has long been a standard control in language model interfaces, governing the randomness of token sampling: lower values push the model toward more deterministic, focused output, while higher values increase variability and can encourage more diverse or creative responses. Removing it means that code paths that set temperature explicitly will need to be updated, and teams that relied on tuning it to stabilize output or to add variety will have to find alternative approaches. Depending on how Anthropic has implemented the change, developers may need to lean on other sampling controls or on prompt design to achieve similar effects. The move appears to reflect a broader trend toward simplifying model interfaces, though it also reduces one lever that some applications used for fine-grained control.

The pricing change is likely to draw the most scrutiny. An effective increase of around 30 percent affects the unit economics of any product built on the model, and the impact is amplified for high-volume applications or for those that routinely send large prompts. Because token-based billing scales with both input and output length, the combination of a much larger context window and higher per-token costs means that careless use of the full one-million-token capacity could raise bills substantially. Teams that adopt Sonnet 5 will need to weigh the benefits of the expanded context and any quality improvements against the higher running costs, and may respond by trimming prompts, caching where possible, or routing lighter tasks to smaller models.

For context, these changes land in a competitive environment where rival providers have pushed in similar directions. Google's Gemini models have offered context windows of a million tokens or more, and OpenAI has continued to iterate on its own model families and pricing. Large context windows have become a common point of comparison, alongside latency, tool use, and reasoning ability. Anthropic has also invested in surrounding capabilities such as prompt caching, batch processing, and its Model Context Protocol, an open standard for connecting models to external data sources and tools, all of which shape the practical cost and performance of building on Claude.

Developers evaluating the upgrade should approach it as both a capability gain and a migration task. The larger standardized context and any underlying model improvements may justify the switch for workloads that benefit from long inputs, but the removal of temperature and the higher prices mean existing integrations cannot simply be pointed at the new model without review. As with previous Claude releases, the real-world tradeoffs will become clearer as teams run their own benchmarks and monitor spending over time. Anyone depending on the specifics should confirm the details against Anthropic's official documentation, since figures and parameter behavior can be adjusted after launch.