ModelsClaude Opus 4.5: The Release That Cut Opus Pricing by Two-Thirds
Opus 4.5 dropped Opus pricing from $15/$75 to $5/$25 and introduced the effort parameter. What it still does well, its 200K limits, and the retirement date ahead.
Most model releases are remembered for a benchmark. Claude Opus 4.5 is better remembered for two decisions that outlasted it: it cut the price of the Opus tier to a third, and it introduced the effort parameter that every later Claude model uses to trade depth for cost.
It is still served nearly a year later. Whether you should still be running it comes down to its 200K context window and a retirement date that is now close.
The price change, in numbers
Anthropic released Claude Opus 4.5 on 24 November 2025 at $5 per million input tokens and $25 per million output. Its predecessor, Claude Opus 4.1, was priced at $15 and $75. Cache reads dropped from $1.50 to $0.50 per million tokens, and batch pricing to $2.50 and $12.50. The catalogue's :batch entry is that same model on the half-price asynchronous endpoint.
The significance was not only the discount. Before Opus 4.5, the top tier was something many teams reserved for the hardest prompts and routed everything else away from. At $5 and $25 it became plausible as a default for agent loops, and every Opus release since has held that price or gone lower.
Effort: the parameter that arrived with 4.5
Opus 4.5 introduced effort on the Claude API, letting a caller trade thoroughness against speed and token spend. On this model it accepts low, medium and high, with high as the default. The xhigh and max levels came later.
Anthropic's own framing of the parameter was token efficiency. It reported that at medium effort Opus 4.5 matched Claude Sonnet 4.5's best SWE-bench Verified score while using 76% fewer output tokens, and that at its highest effort it exceeded Sonnet 4.5 by 4.3 percentage points while using 48% fewer tokens.
Read those claims precisely. They compare output tokens on one benchmark against a different model, not dollar cost on your workload. Output tokens are the expensive side of the bill, so a large reduction there matters, but reasoning-heavy models can still surprise you with hidden thinking spend. The token cost of reasoning models explains where that spend goes.
One detail matters for migrations: Opus 4.5 uses effort alongside the older manual thinking mode, set with budget_tokens. Adaptive thinking, where the model decides how much to think, arrived with Opus 4.6.
200K context and 64K output in a 1M world
Opus 4.5 has a 200K-token context window and 64K maximum output tokens, with text and image input. Anthropic lists its reliable knowledge cutoff as May 2025.
In late 2025 that was a normal specification. By late 2026, 1M-token windows are standard across the current Claude line and much of the open-weight field. For single-file work and most chat, 200K is plenty. For whole-repository reasoning, long agent transcripts or large document sets, it is a wall, and you either retrieve selectively or move to a model with a larger window.
Opus 4.5 or Opus 4.6?
The nearest alternative is its direct successor, and the comparison is lopsided. Claude Opus 4.6 has the same $5 and $25 price, a 1M-token window, 128K output, adaptive thinking and an extra max effort level. It also uses the same older tokenizer, so token counts move across almost unchanged.
Reasons to stay on 4.5 are therefore mostly about stability rather than capability:
- You have evaluations, prompts and output parsers validated against 4.5 and no budget to re-validate before its retirement.
- You depend on assistant prefill, which Opus 4.5 still accepts and the 4.6 generation rejects.
- Your requests set explicit thinking budgets and you have not yet tuned an effort level to replace them.
If none of those apply, Opus 4.6 is the same price with more headroom, and newer Opus models are worth testing too.
The retirement date to plan around
Anthropic lists Opus 4.5 as an active legacy model with retirement not sooner than 24 November 2026. That is about five weeks from the date of writing. Not sooner than is a floor, not a schedule. At the time of writing Anthropic's deprecation page shows no deprecation notice for it, and Anthropic commits to at least 60 days' notice before retiring a public model, so it will not vanish on that date. What the date does tell you is that a deprecation notice can arrive at any point, and that new work should not start here.
Bedrock and Google Cloud set their own lifecycle dates. If you call Opus 4.5 through a partner cloud, check its schedule separately.
How to measure whether effort saves you money
Whether you stay on 4.5 for now or move to a newer model, the effort parameter is the cheapest lever you have. A quick way to find the right level:
- Take twenty representative tasks from real traffic, with a way to grade each one.
- Run them at
low,mediumandhigh, recording pass rate, input tokens, output tokens and turns to completion. - Compute cost per passed task at each level, not cost per request. A cheaper request that needs a retry is not cheaper.
- Pick the lowest level whose pass rate is within your tolerance of the best, and route only the tasks that fail it to a higher level.
Cost-adjusted model scoring covers how to turn those numbers into a ranking you can defend when the budget question comes.
Common questions
How much does Claude Opus 4.5 cost?
Anthropic lists it at $5 per million input tokens and $25 per million output, with cache reads at $0.50 and batch processing at half price. That is a third of the $15 and $75 charged for Opus 4.1.
What effort levels does Claude Opus 4.5 support?
Low, medium and high, with high as the default. The xhigh and max levels arrived on later Opus models, and Opus 4.5 still uses manual budget_tokens for thinking rather than adaptive thinking.
When will Claude Opus 4.5 be retired?
Anthropic commits to not retiring it before 24 November 2026 and gives at least 60 days of notice for public models. Treat late 2026 onwards as the window in which a deprecation notice could arrive.
