On 22 September 2026 Anthropic released Claude Opus 5.5, the first model in the 5.5 family. In Anthropic's own words: "at default settings it will cost 40% less than Opus 5 on typical workloads". The list price dropped too, from $5/$25 to $4/$20 per million input and output tokens. If you run an AI feature, an agent or a coding assistant from Hyderabad, Chennai or Bengaluru, that is a real line item moving. But a 20% price cut is not a 40% smaller bill, and a cheaper frontier model is still wrong for plenty of workloads.
What Anthropic actually shipped on 22 September 2026
Claude Opus 5.5 is positioned for long-running agentic coding and knowledge work, and Anthropic's docs now tell developers to start with it for most workloads. Two things changed at once: the published price per token fell, and the amount of work the model does by default fell with it. Both matter to your bill, and only one shows up on a price sheet.
- •Price: $4 per million input tokens and $20 per million output, down from $5 and $25 on Opus 5 - a 20% cut on both sides of the meter.
- •Cache reads cost $0.20 per million tokens: 5% of base input price rather than the usual 10%, and half the $0.50 Opus 5 charged.
- •Context is 1M tokens with max output of 128K, and the full window is served at standard pricing - long prompts carry no premium rate.
- •Adaptive thinking is always on and cannot be disabled; the effort parameter is now the only control over how much the model reasons.
- •Default effort is medium, where Opus 5 defaulted to high - so an unchanged request runs one level lower than it used to.
- •Claude Sonnet 5.5 and Haiku 5.5 are stated as following "in the coming weeks" - no price, date or benchmark published for either.
Why a 20% price cut turns into a 40% drop in running cost
The two headline numbers are not in conflict, and the gap between them is the whole point. The 20% figure is arithmetic: the published rate per token fell by a fifth. The 40% figure measures typical workloads at default settings - and the defaults changed. Because Opus 5.5 ships with effort set to medium while Opus 5 defaulted to high, a request whose code you never touched now spends fewer tokens thinking and answering. Fewer output tokens at a lower price per token compounds into roughly the 40% Anthropic measured. That also tells you how to lose the saving. Set effort to xhigh or max and you buy back every token the default gave up. The discount is a default, not a guarantee.
Claude Opus 5.5 vs Opus 5 vs Sonnet 5 vs GPT-5.6 Sol
Every cell came from the vendor's own documentation on 24 September 2026, and where a figure is unpublished the table says so rather than guessing. Note the awkward line for anyone treating this as a price war: on list price per token, Claude Opus 5.5 and OpenAI's GPT-5.6 Sol are currently identical.
| Specification | Claude Opus 5.5 | Claude Opus 5 | Claude Sonnet 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Input per million tokens | $4 | $5 | $2 | $4 |
| Output per million tokens | $20 | $25 | $10 | $20 |
| Cached input read | $0.20 | $0.50 | $0.20 | $0.40 |
| Context window | 1M tokens | 1M tokens | 1M tokens | 1,050,000 tokens |
| Max output | 128K tokens | 128K tokens | 128K tokens | 128,000 tokens |
| Thinking mode | Adaptive, always on | Adaptive | Adaptive | not published |
| Default effort | medium | high | high | not published |
| Batch discount | 50% off input and output | 50% off input and output | 50% off input and output | not published |
| Knowledge cutoff | June 2026 | May 2026 | January 2026 | 16 February 2026 |
What does Claude Opus 5.5 actually cost in rupees?
A worked example, built from published list prices, not a client's bill. Suppose a mid-market firm runs a document-analysis agent: 1,500 tasks a month, each sending 40,000 input tokens (a long contract plus instructions) and producing 6,000 output tokens including thinking. That is 60 million input and 9 million output tokens a month, converted at ₹96 per US dollar - rounded from the xe.com mid-market rate of 1 USD = 95.9508 INR quoted at 09:25 UTC on 24 September 2026. Run the same volumes through five options and the spread is far wider than the 20% headline suggests.
- •Opus 5.5 at list: 60 MTok x $4 plus 9 MTok x $20 = $420 a month, about ₹40,320.
- •Opus 5 at list, identical volumes: $300 plus $225 = $525 a month, about ₹50,400. The saving is $105, or exactly 20% - the price cut, nothing more.
- •Opus 5.5 through the Batch API at 50% off, for work that can wait: $210 a month, about ₹20,160.
- •Opus 5.5 with caching, if 30,000 of the 40,000 input tokens per task are cache hits: $60 fresh input plus $9 cache reads plus $180 output = $249, about ₹23,904, before cache writes.
- •Sonnet 5 at the same volumes: $120 plus $90 = $210 a month, about ₹20,160 - half the Opus 5.5 bill, if your evals say quality holds.
- •The failure case: raise effort so output goes from 6,000 to 12,000 tokens per task and Opus 5.5 costs $600 a month, about ₹57,600 - more than Opus 5 charged before the cut.
Which workloads justify Opus 5.5, and which do not
Anthropic's guidance is to start with Opus 5.5 for most workloads and step up to Claude Fable 5.1 only for demanding reasoning and long-horizon agentic work. That is sound advice for a developer picking a default. It is not a procurement policy. The honest test is whether a task's failure is expensive. Where a wrong answer costs you a contract, a customer or a day of engineering time, frontier capability pays for itself. Where a human catches it instantly, you are buying insurance you do not need.
- •Worth it: long-document analysis where the 1M context earns its keep - tender packs, multi-year contracts, whole codebases - all at standard pricing.
- •Worth it: agentic coding and long-running internal tooling, the workloads Anthropic built this model for, where Terminal-Bench 4.0 jumped from 52.3% to 66.4%.
- •Worth it: multi-step agents chaining tool calls, where one weak reasoning step derails the run and you pay for the retry anyway.
- •Not worth it: classification, tagging, routing and extraction - Claude Haiku 4.5 lists at $1/$5 per million tokens, a fifth of the Opus 5.5 rate.
- •Not worth it: high-volume chat and summarisation, where Claude Sonnet 5 at $2/$10 halves the bill and your evals show no measurable quality loss.
- •Not worth it yet: anything you planned to rebuild around Sonnet 5.5 or Haiku 5.5, which have no published price or date to plan against.
Why your AI bill will not fall by 40%
Vendor cost claims describe the vendor's test conditions, not your traffic. Your settings are probably not default, and your bill includes things the benchmark never paid for: failed runs you retry, context you accumulate, documents you paste in because it beat retrieving the right paragraph. These are the five mistakes we see most when a team migrates for the price.
⚠️Treating the 40% figure as a forecast for your own bill
Consequence: Budgets are cut on a number describing Anthropic's test conditions, then blown when real traffic behaves differently.
Solution: Re-baseline cost and latency at your chosen effort level before changing any budget - the step Anthropic's migration guide puts on the checklist.
⚠️Carrying effort settings straight over from Claude Opus 5
Consequence: Requests that omit effort silently drop from high to medium and may lose quality, while requests pinned to xhigh or max erase the saving.
Solution: Run a fresh sweep across low, medium, high, xhigh and max on your own evals, and set the level per workload rather than per organisation.
⚠️Assuming thinking can still be switched off to save tokens
Consequence: Requests setting thinking to disabled return a 400 error, and workloads that previously ran without thinking now emit thinking tokens, billed as output tokens.
Solution: Remove the thinking flags, lower effort where you want less reasoning, and revisit max_tokens - a hard limit covering thinking plus response text.
⚠️Routing every request to the frontier model because it got cheaper
Consequence: Classification and extraction that Claude Haiku 4.5 handles at $1/$5 gets billed at $4/$20, multiplying the bill fivefold on your highest-volume tasks.
Solution: Route by task class, not by habit: a cheap model for structured work, a frontier model for consequential reasoning, with evals deciding the boundary.
⚠️Filling the 1M context window because it costs no extra per token
Consequence: Input tokens dominate the invoice, latency rises and cached prefixes stop matching, so the cheapest tokens become the largest line on the bill.
Solution: Retrieve the relevant passages instead of pasting whole archives, keep a stable cached prefix, and hold effort constant within a cached conversation.
Migration checklist: moving from Opus 5, Fable or GPT
Opus 5.5 is not a drop-in replacement. Anthropic documents four breaking changes against Opus 5, and the ones that bite hardest are the quiet ones - a tool_choice value that now errors, a streamed progress message that goes silent. Work through these in a development environment before production traffic moves.
Swap the model ID first, and nothing else
Point your client at claude-opus-5-5 in a development environment and run your existing suite before changing any other parameter.
Remove the thinking flags and set effort explicitly
Both thinking disabled and thinking enabled are rejected. Pass an explicit effort level instead: the default silently puts you on medium where Opus 5 gave you high.
Fix forced tool use and thinking blocks
Replace tool_choice values of any and tool with auto plus strict tool use, read content blocks by type, and pass thinking blocks back unmodified.
Check what users see between tool calls
Text between tool calls now arrives in thinking blocks that are empty at the default display setting, so an interface streaming progress updates goes quiet unless you opt in.
Run an effort sweep against your own evals
Sweep the effort levels on tasks that represent your real traffic, record quality and token spend at each, and set a large max_tokens at the higher levels.
Re-baseline cost, then move traffic in stages
Measure rupees per completed task at your chosen effort level, shift a slice of production traffic, and widen once retries and refusals hold steady.
Check the commercial small print
Priority Tier is not supported on Opus 5.5, and fast mode is a separately priced preview at $8/$40 per million tokens.
What should you ask an AI vendor before switching models?
A model launch is a good moment to audit whoever builds or runs your AI, because a price change exposes how much they understand about your spend. Ask for numbers tied to your workload and your currency, not a benchmark table. If your vendor cannot answer these in writing, the 40% headline will not reach your bank account.
- •What is our current cost per completed task in rupees, split into input, output, cache reads and cache writes?
- •Which workloads run on which model today, and what evidence decided each of those choices?
- •What effort level does each workload run at, and what did the effort sweep show for quality and token spend?
- •What share of our input tokens are cache hits, and what would it take to raise it?
- •Which jobs could move to the Batch API at 50% off without anyone noticing the delay?
- •What is our retry and refusal rate, and is it in the cost per task you quoted?
How Tech Arion helps with AI model selection and cost control
We build AI features and agents for Indian businesses, and model choice is the part clients get talked into rather than measured into. Our AI consulting practice at techarion.com/services/ai-consulting starts with a cost baseline - what each workload spends per completed task today - then runs an effort and model sweep against evals built from your real traffic, so choosing between Opus 5.5, Sonnet 5 and a cheaper tier is a measured decision, not a vendor headline. Where the work is software, our AI-assisted development team at techarion.com/services/vibe-coding implements the migration properly: explicit effort levels, prompt caching designed into the prefix, batch processing for anything that can wait, and routing so high-volume classification never touches a frontier model. We run the same discipline on our own platform at techarion.com/services/ticket-agent, which is why we quote in rupees per task, not dollars per million tokens.
Frequently asked questions about Claude Opus 5.5 pricing
The questions Indian founders and CTOs have asked us most since the 22 September 2026 release.
Frequently Asked Questions
Work out what Claude Opus 5.5 is actually worth to you
A cheaper frontier model only helps if the right workloads are on it. Tech Arion baselines what your AI features cost per completed task in rupees, builds evals from your real traffic, runs the effort and model sweep, and implements the migration with caching, batching and task routing in place. Talk to us about a cost baseline before you move production traffic to a new model.
Sources & References
Primary sources fetched on 24 September 2026:
- 1.
Anthropic. (22 Sep 2026). Introducing Claude Opus 5.5 - "at default settings it will cost 40% less than Opus 5 on typical workloads", $4/$20 per million tokens, output more than 30% faster, Terminal-Bench 4.0 66.4% vs 52.3%, Sonnet 5.5 and Haiku 5.5 to follow in the coming weeks.
View Source - 2.
Anthropic. (2026). Claude platform pricing - Opus 5.5 $4/$20, cache read $0.20, cache writes $5 and $8; Opus 5 $5/$25; Sonnet 5 $2/$10; Haiku 4.5 $1/$5; 50% Batch API discount; 1M context at standard pricing; fast mode $8/$40.
View Source - 3.
Anthropic. (2026). Claude Opus 5.5 model overview - 1M context, 128K max output (300K on the Batch API beta), adaptive thinking always on, default effort medium, knowledge cutoff June 2026, released 22 September 2026.
View Source - 4.
Anthropic. (2026). Claude Opus 5 model overview - legacy status, released 24 July 2026, $5/$25 per million tokens, cache read $0.50, 1M context, default effort high, knowledge cutoff May 2026.
View Source - 5.
Anthropic. (2026). Claude models overview - current lineup specifications used for the Claude Sonnet 5 and Claude Haiku 4.5 comparison, and the guidance to start with Opus 5.5 for most workloads.
View Source - 6.
Anthropic. (2026). Effort parameter documentation - five effort levels, Opus 5.5 defaults to medium where other models default to high, effort affects all output tokens including thinking, and changing top-level effort invalidates prompt caching.
View Source - 7.
Anthropic. (2026). Migrating to Claude Opus 5.5 - breaking changes and the migration checklist, including re-baselining cost and latency at your chosen effort level and that thinking tokens are billed as output tokens.
View Source - 8.
TechCrunch. (22 Sep 2026). Anthropic releases Opus 5.5 - "Output tokens will be charged at $20 per million tokens for Opus 5.5, compared to $25 for the previous model", and Sonnet 5.5 and Haiku 5.5 due in the coming weeks.
View Source - 9.
MarketScreener. (2026). Anthropic launches Claude Opus 5.5 - $4/$20 per million tokens at "rates that are 20% lower than Opus 5", available on Amazon Web Services, Google Cloud and Microsoft Azure.
View Source - 10.
OpenAI. (2026). GPT-5.6 Sol model documentation - $4 per million input tokens, $0.40 cached input, $20 per million output tokens, 1,050,000 context window, 128,000 max output, knowledge cutoff 16 February 2026.
View Source - 11.
xe.com. (24 Sep 2026). USD to INR mid-market rate - 1 USD = 95.9508 INR at 09:25 UTC, rounded to ₹96 per dollar for the worked cost model in this article.
View Source