If you are checking Claude Opus 4.5 pricing in July 2026, treat the query as a legacy-active model check, not a latest-model buying guide. Anthropic's current pricing page still lists Claude Opus 4.5 at $5 per million base input tokens and $25 per million output tokens, with cheaper cache-hit and batch rows. But Anthropic's model overview now points current Opus work to Claude Opus 4.8 with the API ID claude-opus-4-8.
The practical answer is:
| Situation | Better default |
|---|---|
You already have production traffic pinned to claude-opus-4-5-20251101 | Keep it only while you audit usage, retirement dates, and migration risk. |
| You are starting a new Opus build | Test claude-opus-4-8 first instead of choosing 4.5 from an old article. |
| You only care about price | Do not switch among Opus 4.5, 4.6, 4.7, and 4.8 for a lower standard Opus price; optimize model choice, cache, batch, context, and output length instead. |
| You are comparing API cost with Claude Pro or Max | Use the Claude API pricing vs subscription guide because subscription allocation and API billing are different contracts. |
| You use AWS, Google Cloud, or another cloud route | Check that platform's current price page and region mode; Anthropic documents regional and data-residency premiums for some routes. |
Current Claude Opus 4.5 API Pricing
Anthropic's official pricing page currently lists the Opus 4.5 row this way:
| Token category | Current Opus 4.5 price |
|---|---|
| Base input tokens | $5 / MTok |
| 5-minute cache write | $6.25 / MTok |
| 1-hour cache write | $10 / MTok |
| Cache hit and refresh | $0.50 / MTok |
| Output tokens | $25 / MTok |
| Batch input | $2.50 / MTok |
| Batch output | $12.50 / MTok |
MTok means one million tokens. The important asymmetry is still output cost: output tokens cost five times the base input token rate. A long coding answer, reasoning trace, generated report, or verbose JSON payload can dominate the bill even when the prompt is not large.
For a normal synchronous request, the rough API formula is:
hljs textcost = (input_tokens / 1,000,000 * 5) + (output_tokens / 1,000,000 * 25)
For eligible batch work, the formula changes to:
hljs textbatch_cost = (input_tokens / 1,000,000 * 2.5) + (output_tokens / 1,000,000 * 12.5)
Prompt caching changes only the repeated input side. A 5-minute cache write costs more than normal input on the first write, but a cache hit is priced at 10% of base input. That can matter for repeated system prompts, shared docs, policy context, tool schemas, and large stable instructions.
What Changed Since Opus 4.5 Was the Fresh Model
The old version of this page treated Opus 4.5 as the newest premium Claude route. That is no longer the right framing.
Anthropic's current model overview lists Claude Opus 4.8 as the Opus row for complex agentic coding and enterprise work, with claude-opus-4-8 as the Claude API ID. Anthropic's Opus 4.8 launch note also states regular usage pricing at $5 per million input tokens and $25 per million output tokens.
That means a price-only comparison does not justify staying on Opus 4.5. If you are starting today, the question is whether Opus is needed at all, and if it is, whether Opus 4.8 passes your workflow better than the older pinned model. If you already run 4.5 in production, the decision is an engineering migration question: eval coverage, prompt behavior, token mapping, tool behavior, cloud route, latency, and fallback plan.
The lifecycle boundary matters. Anthropic's model deprecations page currently lists claude-opus-4-5-20251101 as active, with a tentative retirement date not sooner than November 24, 2026. Active does not mean future-proof. It means existing users still have time to audit and migrate deliberately.
Should You Keep Using Claude Opus 4.5?
Keep Opus 4.5 temporarily when the model ID is already pinned and the cost of a rushed migration is higher than the benefit of moving today. Typical cases include:
- production evals are tuned to
claude-opus-4-5-20251101; - a cloud provider or internal gateway has not yet exposed the newer route you need;
- output format stability matters more than trying the newest model this week;
- your team needs to compare the same workload across 4.5 and 4.8 before changing defaults.
Move the default test path to Opus 4.8 when:
- the workload is new;
- the old model ID appears only because an article, template, or config was copied;
- you are doing coding-agent, enterprise, or long-running workflow tests;
- you need a model with a longer expected lifecycle;
- you have enough eval coverage to compare output quality and token usage on the same prompts.
Do not keep 4.5 because of vague claims like "more stable", "cheaper", or "better for coding" unless your own logs prove that for your workload. The official standard Opus price row is not lower for 4.5, and model behavior changes are task-specific.
Cost Controls That Still Matter
The strongest cost levers are not keyword-friendly discount claims. They are the boring controls that change billable tokens.
| Lever | When it helps | Watch out for |
|---|---|---|
| Smaller model routing | Classification, extraction, light summarization, routine rewrites | Do not route hard tasks to a cheaper model if retries erase the savings. |
| Prompt caching | Repeated stable context across requests | Cache writes have an upfront cost; rapidly changing prompts reduce the benefit. |
| Batch API | Non-urgent offline jobs, document runs, overnight processing | Not suitable for interactive chat or real-time agents. |
| Output caps | Code generation, JSON, reports, long explanations | Too-low caps can create partial outputs and retries. |
| Context trimming | Large repos, long docs, long Claude Code sessions | Removing the wrong context can hurt quality. |
| Spend and rate limits | New agents, queues, CI jobs, customer-facing tools | Limits protect budgets but can create 429s if capacity planning is weak. |
For a page about Opus 4.5 pricing, this is the safe conclusion: price the old model accurately, but optimize the workload rather than pretending an old model ID is a discount strategy.
API, Subscription, and Cloud Route Boundaries
Claude API pricing is not the same thing as a Claude Pro or Max subscription. A subscription is a human-product allocation for Claude surfaces. API pricing is token-metered developer usage through Console or supported cloud routes. Claude Code can sit on either side depending on authentication and environment, so check /status and API keys before comparing costs.
Cloud routes also need their own check. Anthropic's pricing documentation says Claude models are available through Amazon Bedrock and Google Cloud, and that regional or multi-region endpoint choices can add premiums for Claude 4.5 models and later. Do not copy a first-party Claude API number into a cloud budget without checking the cloud provider's current regional price and billing unit.
This page also no longer recommends a fixed third-party discount. Any aggregator or gateway claim about lower price, latency, availability, support, refund, uptime, or model coverage is freshness-sensitive. Verify that route directly on the same day, and avoid sending sensitive or regulated workloads through a provider whose data handling and support contract you have not reviewed.
Migration Checklist
Before replacing claude-opus-4-5-20251101 with claude-opus-4-8, run the migration as a measured test:
- Export a representative sample of real prompts, tool calls, files, and expected outputs.
- Run the same workload on Opus 4.5 and Opus 4.8.
- Compare task success, output length, latency, tool-call shape, and retry rate.
- Recalculate cost from actual input, cache, and output tokens rather than list price alone.
- Check cloud route availability and any regional premium.
- Keep a rollback path until the new default has passed real production-like cases.
- Recheck Anthropic's deprecation page before each quarterly planning cycle.
If the workload is mostly "which Claude model should I choose?", the better next read is Opus vs Sonnet. If the workload is "will my subscription cover this?", use Claude API pricing vs subscription.
FAQ
Is Claude Opus 4.5 still priced at $5 input and $25 output per million tokens?
Yes, Anthropic's current pricing page still lists Claude Opus 4.5 at $5 per million base input tokens and $25 per million output tokens. Recheck the official pricing page before publishing budgets or customer-facing quotes.
Is Claude Opus 4.5 still the latest Opus model?
No. Anthropic's current model overview lists Claude Opus 4.8 as the Opus route for complex agentic coding and enterprise work. Use Opus 4.5 only when you have a specific compatibility reason.
Is Claude Opus 4.5 retired?
No. Anthropic's model deprecations page currently lists claude-opus-4-5-20251101 as active, with a tentative retirement date not sooner than November 24, 2026. That is enough time to migrate carefully, not a reason to ignore lifecycle risk.
Does prompt caching really cut Opus 4.5 cost by 90%?
Only for repeated input that becomes a cache hit. Anthropic prices cache hits at 10% of base input, but outputs still bill at the normal output rate unless another eligible modifier applies.
Is batch cheaper for Opus 4.5?
Yes, Anthropic lists batch input at $2.50 per MTok and batch output at $12.50 per MTok for Opus 4.5. Use it only for asynchronous work that can wait.
Should I choose Opus 4.5 instead of Opus 4.8 to save money?
Usually no. The standard Opus price row is the same for Opus 4.5 and Opus 4.8 in Anthropic's current pricing docs. Choose based on eval results, lifecycle, route availability, and workload behavior.
Can a third-party gateway make Opus 4.5 cheaper?
Possibly, but this page no longer treats any fixed unofficial discount as a publishable fact. Verify current model availability, price, latency, data handling, support, and billing records directly before routing production traffic.
