Home / AI

Claude Opus 4.5 Clears 80% on SWE-bench Verified and Cuts Opus Pricing by Two-Thirds

Anthropic's new flagship leads coding benchmarks over Gemini 3 Pro and GPT-5.1-Codex-Max while dropping to $5 and $25 per million tokens.

AINov 24, 20254 min read2 diagrams

Key takeaways

  • Opus 4.5 scored 80.9% on SWE-bench Verified, ahead of GPT-5.1-Codex-Max (77.9%) and Gemini 3 Pro (76.2%).
  • API price fell to $5 input and $25 output per million tokens, from $15 and $75 for Opus 4.
  • It launched November 24, 2025 across Claude apps, Claude Code and the API.

Anthropic released Claude Opus 4.5 on November 24, 2025, the third frontier model to land in under two weeks. It is the first to report more than 80% on SWE-bench Verified, and it costs a third of what earlier Opus models did.

A coding lead, narrowly

SWE-bench Verified gives a model real GitHub issues and asks it to produce a patch that passes the project's tests. Opus 4.5 scored 80.9%, according to figures reported across several outlets. OpenAI's GPT-5.1-Codex-Max came in at 77.9% and Google's Gemini 3 Pro at 76.2%. The gaps are a few points, not a chasm, and each lab runs its own scaffolding, so the ranking is best read as a statement of who is currently ahead rather than by how much.

Bar chart of SWE-bench Verified: Claude Opus 4.5 80.9, GPT-5.1-Codex-Max 77.9, Gemini 3 Pro 76.2 SWE-bench Verified (%) Claude Opus 4.5 80.9 GPT-5.1-Codex-Max 77.9 Gemini 3 Pro 76.2 Bars start at zero; differences are small in absolute terms.
SWE-bench Verified scores as reported at launch. Source: Neowin, OfficeChai, Digit.

The price cut is the bigger story

Opus 4 and 4.1 listed at $15 per million input tokens and $75 per million output. Opus 4.5 is priced at $5 and $25. That is a 67% reduction, and it moves the top-tier model into the same band as mid-range competitors: GPT-5.1 was $1.25 and $10, Gemini 3 Pro $2 and $12, so Opus remains the premium option but no longer an outlier. For agentic coding, where a single task can chew through hundreds of thousands of tokens, the rate card changes which jobs are worth automating.

Grouped bars comparing per-million-token prices: Opus 4 at 15 input and 75 output, Opus 4.5 at 5 input and 25 output Opus API price, USD per million tokens $15 $75 $5 $25 Opus 4 input output Opus 4.5 input output -67%
Anthropic list prices for Opus 4 versus Opus 4.5. Source: Free Press Journal, Neowin.

Why it matters

Three frontier releases in two weeks (GPT-5.1 on November 12, Gemini 3 Pro on November 18, Opus 4.5 on November 24) turned the autumn into a cadence race. Each lab now ships a clear lead on one metric and loses it within days. What persists is the price curve: capability at a given tier gets cheaper every quarter, and Anthropic chose to compete on it directly.

Coding remains the commercial anchor. Anthropic's revenue leans heavily on developer tools such as Claude Code, so a model tuned for long, multi-step software tasks, available in the same apps and API on day one, protects that position against Google's Antigravity and OpenAI's Codex line.

What the benchmark does and does not say

SWE-bench Verified is a curated subset of real issues from popular open-source Python projects, checked by humans to confirm that each task is solvable and its tests fair. Passing means the model's patch makes the failing tests pass without breaking others. That is a tougher standard than writing a function from a docstring, but it is still narrow: the repositories are public and well-known, so some leakage into training data is plausible, and the tasks are mostly bug fixes rather than the open-ended feature work developers actually assign. A model at 80.9% does not resolve four out of five tickets in your codebase.

That is why the surrounding claims matter. Anthropic positioned Opus 4.5 as the best model for coding and agentic tasks generally, and outlets noted that it also beat Gemini 3 Pro on agentic-use benchmarks. The model shipped in the Claude apps, in Claude Code, and through the API on day one, so customers did not wait for a staggered rollout. Coverage also pointed to new controls for developers, including more control over how much effort the model spends on a task, which lets teams trade cost and latency against thoroughness on a per-request basis.

The arithmetic of the price cut deserves a worked example. A long agentic session that consumes two million input tokens and a quarter of a million output tokens cost about $48.75 at Opus 4 prices ($30 plus $18.75). At Opus 4.5 prices the same session costs $16.25 ($10 plus $6.25). Multiply that by hundreds of sessions a day inside a software team and the difference is a line item finance notices. Lower prices also change behavior: teams that rationed the flagship model for hard problems can use it as the default.

The competitive response will reveal how much room remains. Google has a lower per-token list price on Gemini 3 Pro, and OpenAI's coding line is tuned for sustained autonomous runs. Anthropic is wagering that quality on long, tool-heavy tasks, rather than the lowest price, wins developers, and that a smaller gap in cost removes the usual reason to trade down.

What to watch

  • Independent runs of SWE-bench and Terminal-Bench with matched scaffolding.
  • Whether OpenAI answers with a flagship refresh; it did so within weeks.
  • How Opus-level pricing pressure shows up in Cursor, Copilot and other resellers' plans.

Sources

  1. Anthropic launches Claude Opus 4.5, claiming it's the world's best model for coding — Neowin
  2. Anthropic Releases Claude Opus 4.5, Beats Gemini 3.0 On Coding, Agentic Use Benchmarks — OfficeChai
  3. Anthropic introduces Claude Opus 4.5, its most capable model yet — Digit
  4. Anthropic Rolls Out Claude Opus 4.5: Brings New Abilities, Reduced Prices — Free Press Journal

More in AI