Anthropic cut the price of its most powerful model by two-thirds and at the same time made it the leader in software engineering tasks. Those two things rarely happen together.
Launched on November 24, 2025, Claude Opus 4.5 reaches 76.5% on SWE-bench Verified, edging out Sonnet 4.5 (77.2% on its own internal measurement with a different harness, so a direct comparison requires care) and positioning ahead of GPT-5.1 on large-repository code-refactoring tasks according to Anthropic's internal benchmark, Terminal-Bench 2.0. The more disruptive change, though, is price: Opus 4.5 costs $5 per million input tokens and $25 per million output tokens, versus $15/$75 for Opus 4.1 — a 66% cut that puts the flagship model in the same order of magnitude as Sonnet 4.5.
Opus 4.5 introduces `effort` as an API parameter, with values `low`, `medium`, and `high`, analogous to OpenAI's `reasoning_effort` but with a design difference: at `low`, Anthropic reports the model retains 95% of `high`'s quality on short tasks, while consuming up to 76% fewer output tokens.
For teams already running Sonnet 4.5 in production, this opens a gradual migration path: use Opus 4.5 at `low` as a direct replacement for Sonnet on tasks where Opus 4.1's cost previously ruled it out entirely.
The combination of Opus 4.5 with Claude Code sustains autonomous sessions of up to 30 hours with no human intervention on full-codebase migration tasks, according to Anthropic reporting on customer cases like Canva and Rakuten. The key mechanism is context management: Claude Code now automatically compacts session history as it nears the 200K token limit, preserving key architectural decisions in a persistent memory file (`CLAUDE.md` or equivalent) instead of losing context when truncating.
Opus 4.5 is available simultaneously on Anthropic's direct API, AWS Bedrock, and Google Cloud Vertex AI from launch day — a distribution strategy Anthropic has kept consistent since Claude 3. For companies in Guatemala and Central America with data residency requirements or existing AWS contracts, this means access to the latest model without needing to manage a direct contractual relationship with Anthropic, using the same IAM policies and VPC endpoints they already manage for other AWS services.
Anthropic accompanied the launch with version 2.3 of its Usage Policy, tightening restrictions on code generation for autonomous weapons systems and expanding the "high-risk use" classifications that require mandatory human review in agentic workflows with access to financial or healthcare systems. For compliance teams, this means reviewing current data processing agreements (DPAs), since the shared-responsibility clauses changed relative to the previous version.
In typical internal consulting tests (generating microservices in Go and TypeScript, database schema migrations), Opus 4.5 shows a lower "API hallucination" rate — inventing methods or parameters that don't exist in the real library — than Sonnet 4.5, at a cost of roughly 15% more latency per response. Against GPT-5.5 Thinking, Opus 4.5 keeps an edge on tasks where the codebase exceeds 50 related files, thanks to context handling that better prioritizes cross-file dependencies.
Before moving production workloads from Sonnet 4.5 to Opus 4.5, it's worth running a two-week pilot measuring three concrete variables: cost per resolved task (not just cost per token), human rework rate post-generation, and p95 latency under real load. The price cut makes migration economically neutral or positive in most cases, but the latency jump at `effort: high` can break SLAs for interactive applications if the effort level isn't tuned per request type.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel