Anthropic Launches Opus 5.5 with 40% Price Reduction
Anthropic has released Claude Opus 5.5, the first model in the Claude 5.5 series, claiming that its performance on most tasks is comparable to Claude Fable 5.1, with typical workload costs about 40% lower than Opus 5.
The pricing is set at $4 per million tokens for input and $20 for output, a 20% reduction from Opus 5's $5 and $25. Cache reads are priced at $0.20, down 60% from Opus 5's $0.50, which the company states is a major cost for agents and programming tasks. Cache writes are $5, compared to the previous generation's $6.25. Under default settings, output speed is over 30% faster. Fast mode can achieve up to about 2.5 times the speed in Claude Code and the platform.
The company claims that under default effort, FrontierCode outperforms GPT-6 Astra with a single-task cost of about 20%; Terminal Bench 4.0 is on par with Astra, with costs around 40%; CursorBench scores 11 points higher than GPT-5.6 Sol, with costs about one-third. An early tester completed around 680,000 lines of code migration in less than a day. The context window is 1 million tokens, with a maximum output of 128,000 tokens, knowledge cutoff in June 2026, and model ID is claude-opus-5-5.
Usage limits for Pro, Max, and Team subscriptions have been increased to five hours, and subscribers can reset their saved rate limits. The company states this is the first release following the call to "slow down the frontier," with behavior audit scores higher than existing models, more resistant to prompt injection, and fewer instances of overreach or irreversible actions. Thinking is enabled by default and cannot be turned off, and forced tool calls will result in errors.
Compared to Fable 5.1's $10 input and $50 output, Opus 5.5 is priced lower. Anthropic positions it for long-term agent programming and knowledge work, rather than the high-end Mythos/Fable. It is available through API, Amazon Bedrock, and Google Cloud.
This is a share grab driven by product price reductions. Buyers are developers and enterprises migrating agent and programming work from Fable or competitors to lower-priced options, while sellers are using the old Opus 5 or Fable pricing. Funds are flowing from high-priced frontier tiers to programming workflows with high cache hit rates; beneficiaries are agent coding clients billed by token, while those under pressure are still selling similar tasks at high prices with Fable and need to adapt code to the unturnable thinking and changing tool call rules.
Source: Public Information
ABAB AI Insight
On July 24, Anthropic positioned Opus 5 as "close to Fable, at half price," and less than two months later, it released 5.5, cutting input and output by 20% and cache reads by 60%. Fable and Mythos remain at the $10/$50 tier, while Opus is being written as a daily agent engine. The company also raised the five-hour subscription limit, effectively exchanging reasoning efficiency for usage, rather than selling the flagship separately at a higher score. The 680,000 line migration case targets clients looking to switch from person-month engineering to token billing.
The capital path is a return of computational efficiency to pricing. The computational power required for Opus 5.5 is lower than for Opus 5, with the price difference allocated to programming loops with high cache hits. Agents repeatedly read the same repositories and tool results, and cache prices can determine total costs more than input and output. Fast mode at 2.5 times speed, double the price, also turns latency into a sellable tier. The listing on Bedrock and Google Cloud spreads the price reduction instantly into enterprise procurement contracts.
It benchmarks against OpenAI's GPT-6 Astra, GPT-5.6 Sol, and previously Kimi K3, which approached Fable with lower output prices. Cursor has already been positioned as a default coding model close to Fable during the Opus 5 phase. The industry has shifted from competing for higher scores to competing for lower dollar amounts for equivalent agent tasks. Anthropic is in a control period where it swallows flagship daily usage with mid-tier pricing.
The structural change is a transfer of pricing power. Intelligence is no longer sold solely by model name but by "dollars per task." The mechanism is that cache and token efficiency convert fixed intelligence into variable costs; whoever first reduces cache read prices to $0.20 will take the warehouse-level agent loops. Flagship high prices are reserved for occasional challenges, while daily coding is taken by mid-tier pricing.
ABAB News · Cognitive Laws
Agent billing is in cache, not on the leaderboard
After the appearance of half-price flagships, the original flagship becomes occasional use
First reduce service computing power, then lower prices, and market share will move