SAN FRANCISCO, Jan. 21 — The default setting is the headline. When Anthropic put Claude Opus 4 on sale Tuesday, it did more than ship a larger model: it decided that the model should think at length before replying unless a developer explicitly opts out. Paid subscribers on Claude.ai and API customers who call the Opus tier get that behavior from the first request.

Opus 4 is the biggest model Anthropic has released since the first Opus 3 in early 2024, and its first frontier-class launch since Sonnet 4.5 late last year. It lands about six weeks ahead of the date OpenAI has widely been expected to ship GPT-5 (our coverage of that launch, when it came in March, is here).

The release at a glance

  • Reasoning: extended thinking is on by default; API callers can switch it off request by request.
  • Benchmark claim: a 79.4 percent solve rate on SWE-bench Verified, the test of whether a model can fix genuine bugs pulled from public code repositories on its own. Anthropic says Opus 4 also sets state-of-the-art marks on other agentic-coding and tool-use evaluations.
  • API price: $15 per million input tokens and $75 per million output tokens — unchanged from Opus 3. Tokens spent thinking are billed as ordinary output.
  • Rate limits: default limits for Opus-tier API accounts are doubled.
  • Context: 200,000 tokens on the standard API, as before; enterprise customers can try a 1-million-token window in beta.
  • Consumer plan: Claude Pro stays at $20 a month, with Opus 4 use governed by an allowance.
  • Where to get it: Claude.ai, the Anthropic API, Amazon Bedrock and Google Cloud Vertex AI, all from launch day.

Why thinking-first is the bet

Extended thinking is not new to Claude. It first appeared as an opt-in mode on Sonnet 3.7 and was reworked for Sonnet 4.5. The mechanism gives the model extra inference budget to plan and reason through a problem before it commits to a final answer.

What changed with Opus 4 is the tuning. Anthropic says the latency cost on everyday questions is now “negligible,” while the payoff on difficult reasoning and coding work is “substantial” — which is the argument for making it the default rather than a switch most users never find. A company spokesperson framed the release in exactly those terms in a written statement:

“Opus 4 is the model we have been building toward since we introduced the Sonnet/Haiku/Opus tiering in early 2024,” the spokesperson said. “It is meaningfully more capable than any previous Claude on long-horizon agentic work, and the extended-thinking default is part of why.”

The billing choice matters for developers. Because thinking tokens are charged at the regular output rate, a request that reasons longer simply costs more, and the per-request off switch is the cost control. Teams running high-volume, simple prompts through Opus will want to test both settings before assuming the new default suits their workload.

What Pro subscribers get

For individuals, the price of Claude Pro does not move. Opus 4 sits behind a usage allowance, and Anthropic described that allowance as “more generous than what Pro users had on Opus 3.” The company did not publish the allowance as a number. Readers weighing a paid chatbot plan against the alternatives can find our framework in How to Pick an AI Chatbot for Daily Work.

A crowded week

The launch shared a news cycle with two rivals’ moves. Mistral AI announced a $1.3 billion funding round the same day, and a week earlier Google had previewed its upcoming Gemini 3 model at an enterprise event in London; Google released Gemini 3 in March.

Beatrix Mendelsohn-Liu, an analyst at the foundation-model research firm Galileo Insights, said the coding numbers were the part to watch. “What is interesting about the Opus 4 release is the agentic coding numbers,” she said. “If the SWE-bench number holds up in independent evaluation, that is a meaningful gap on the most economically valuable benchmark in the model market right now.”

That condition — independent confirmation of a vendor-reported score — is the open question this launch leaves behind. For a head-to-head read once both flagships were in market, see Is ChatGPT-5 Actually Better Than Claude Opus 4?

Ronan Whitfield-Asari covers model launches and developer platforms from Consumer Tech Wire’s San Francisco bureau.