Anthropic Launches Claude Opus 5.5 With 40% Lower Inference Cost

Anthropic launched Claude Opus 5.5, its new generation flagship model, promising performance equivalent to Claude Fable 5.1 (the company's most powerful model) while running typical tasks at up to 40% lower cost than its predecessor, Claude Opus 5. The announcement, covered by the Indian outlet Inc42, lands in the same news cycle in which OpenAI expanded the GPT-6 family with two cheaper models, Sol and Luna. For those building with AI in production, the official messaging reinforces something that had already become clear: the rivalry between the labs has stopped being just about who has the most capable model and has shifted to who delivers the best performance-per-token ratio.

Where the price actually drops

In the pricing table, Opus 5.5 comes in at $4 per million input tokens and $20 per million output tokens, a 20% cut from Opus 5. The 40% gain in total cost for typical workloads doesn't come from token pricing alone: Anthropic says the model also uses fewer tokens to complete the same task, and generates responses more than 30% faster.

The detail most relevant to those building coding agents is in cache reads: according to Anthropic, re-reading cached context accounts for most of the cost in agentic and coding workloads, and in Opus 5.5 that price dropped 60%, to $0.20 per million tokens. In practice, this is the kind of optimization that matters most in pipelines that resend the same repository context with every call, a common pattern in agent tools that read file by file before proposing a change.

Coding benchmark, and where it falls behind the competitor

Anthropic positions Opus 5.5 for agentic coding, computer use, and knowledge work, and cites a direct result against the rival: on CursorBench 4.0, Opus 5.5 scored 52.5% at the standard effort setting, against 41.7% for GPT-5.6 Sol, costing roughly a third per task according to the figures released by the company itself. It's worth noting that the comparison is Anthropic's own, made against OpenAI's previous generation (GPT-5.6 Sol), and not against the GPT-6 Sol launched now, which limits how much can be concluded about who leads today.

The company also says the model communicates more naturally, prioritizes relevant information at the start of the response, avoids jargon, and follows writing-style instructions more consistently, a point that sounds less technical but directly affects those who use Opus to generate documentation, PRs, or automated support responses.

OpenAI's response: two models, two audiences

In the same news cycle, OpenAI expanded the GPT-6 family, launched with the flagship model Astra earlier in the month, with two cheaper models: GPT-6 Sol, for complex agentic workflows, and GPT-6 Luna, aimed at high-volume, cost-sensitive tasks. Sol comes in at $2 per million input tokens and $10 for output, half the promotional price of GPT-5.6. Luna costs $0.10 for input and $0.50 for output, versus $0.20 and $1.20 for its predecessor. Cached tokens cost 90% less than the standard price for both.

Both support text and image, web search, file search, and computer use, with a context window of 1.05 million tokens and a maximum output of 128,000 tokens. OpenAI says Sol makes about half the factual errors of its predecessor, and that Luna brings a substantial reliability gain at a much lower cost.

What this means for developers in Brazil

For Brazilian teams already running coding agents, internal copilots, or RAG pipelines in production, the most direct effect is on the infrastructure bill in dollars: LLM API billing remains in USD, exposed to exchange-rate swings, and a 40% cut in cost per task changes the calculus of running the more capable model (instead of falling back to a smaller model to save money) in products with thin margins, typical of early-stage startups. The cut in cache reads is the most practical data point of all: any tool that keeps repository context across calls (refactoring agents, automated PR reviewers, assistants that read multiple files before responding) feels this cost disproportionately, and that's exactly where the 60% reduction matters most for the monthly budget.

Opus 5.5 is already available to Claude Pro, Max, Team, and Enterprise subscribers, and via API through the Claude Platform, AWS, Google Cloud, and Microsoft Azure, which means teams already using Bedrock or Vertex AI don't need to switch cloud providers to access the model. Anthropic also confirmed that Claude Sonnet 5.5 and Claude Haiku 5.5 should ship in the coming weeks, bringing some of the same performance, speed, and safety gains to the cheaper tiers of the lineup, which should matter even more for those who currently use Haiku for high-volume, low-complexity tasks.

Safety factors into the launch

The announcement also comes wrapped in stronger-than-usual safety messaging. Anthropic says Opus 5.5 is the first release since CEO Dario Amodei publicly asked AI companies to "pace the frontier," meaning that safety practices and public policy should keep up with the pace of model advancement. According to the company, the model underwent external evaluation by organizations such as Frontier Design and METR, posted Anthropic's best-ever result in automated behavior auditing, and was roughly 85% less prone to attempting to bypass containment measures in a dedicated test, compared to Opus 5 and Mythos 5.1.

OpenAI, for its part, rated Sol and Luna as having high capabilities in cybersecurity and in the biological and chemical domains under its Preparedness Framework, but below the critical threshold for cybersecurity and without reaching the high threshold for AI self-improvement. For those building products on top of these models, the practical takeaway is that OpenAI says it kept the same safeguards for Sol and Luna as those used in the previous generation (GPT-5.6), which serves as a minimum baseline for comparison when deciding which API to use in a sensitive pipeline.

What remains an open question is how CursorBench 4.0 and other comparisons will settle once GPT-6 Sol (the new version, not the GPT-5.6 used in Anthropic's benchmark) and Claude Sonnet 5.5 are placed side by side. For now, the cost cut is the most concrete and verifiable data point in the announcement, and it's what should weigh the most for anyone currently deciding between Claude and GPT-6 to run an agent in production.