Anthropic launches Claude Opus 5.5 with lower pricing and stronger coding and agent performance
Anthropic has launched Claude Opus 5.5, the first model in its new Claude 5.5 generation, bringing Opus-class pricing down while pushing the model closer to the capability level of Claude Fable 5.1 across coding, agents, and professional knowledge work.
The new model costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Claude Opus 5. That is a 20% reduction in the headline token price, while Anthropic says the combination of lower prices and improved token efficiency reduces the cost of typical workloads by approximately 40%.
Opus 5.5 also generates output more than 30% faster than Opus 5, according to Anthropic, and uses fewer tokens to complete many tasks. The company is positioning the model around agentic coding and real-world professional work rather than treating the release primarily as a conventional reasoning upgrade.
........
CLAUDE MODEL | INPUT / 1M TOKENS | OUTPUT / 1M TOKENS | POSITIONING |
Claude Opus 5 | $5 | $25 | Previous Opus generation |
Claude Opus 5.5 | $4 | $20 | High-end coding, agents and knowledge work |
Claude Fable 5.1 | $10 | $50 | Frontier Claude tier |
........
··········
OPUS 5.5 CUTS TOKEN PRICES BY 20%
The most immediate change is economic.
Claude Opus 5 launched at $5 per million input tokens and $25 per million output tokens. Opus 5.5 reduces both rates by one fifth, to $4 and $20.
The reduction becomes substantial in workloads that generate large volumes of output, including autonomous coding agents, research systems, document production, and other multi-step workflows where the model remains active for an extended period.
For a simplified workload containing 10 million uncached input tokens and 2 million output tokens:
........
WORKLOAD | OPUS 5 | OPUS 5.5 |
10M input tokens | $50 | $40 |
2M output tokens | $50 | $40 |
Total | $100 | $80 |
........
Data Studios calculation: the same raw token volume costs 20% less on Opus 5.5 before considering any reduction in the number of tokens required to finish the task.
Anthropic's larger claim is that typical workloads can cost approximately 40% less to run because Opus 5.5 combines the lower tariff with greater execution efficiency.
Those are two different measurements. The API price falls by 20%; the claimed workload-level saving incorporates how the model behaves while completing the task.
··········
ANTHROPIC IS PUSHING OPUS CLOSER TO FABLE-LEVEL PERFORMANCE
Anthropic describes Opus 5.5 as performing at approximately the level of Claude Fable 5.1 for most tasks, despite sitting in a considerably cheaper pricing tier.
The company reports that Opus 5.5 exceeds both Opus 5 and Fable 5.1 on most of the evaluations included in its launch material, with particularly strong results in agentic software engineering and practical knowledge work.
This changes the role of the Opus tier inside the Claude model family.
Fable remains the premium frontier option, but its standard API prices of $10 per million input tokens and $50 per million output tokens are 2.5 times the corresponding Opus 5.5 rates.
........
PRICE RATIO | OPUS 5.5 | FABLE 5.1 |
Input | $4 | $10 |
Output | $20 | $50 |
Fable / Opus 5.5 | — | 2.5× |
........
If the two models perform similarly on a given workload, the economics favor Opus 5.5 heavily. The remaining justification for selecting Fable 5.1 therefore depends on workloads where its additional capability produces enough improvement in completion quality or reliability to offset the higher inference cost.
That makes model routing more important than simply selecting the most capable model available.
··········
AGENTIC CODING IS ONE OF THE MAIN AREAS OF IMPROVEMENT
Opus 5.5 continues Anthropic's concentration on software-engineering agents rather than isolated code generation.
The company reports stronger performance on agentic coding evaluations and says the model completes tasks using fewer tokens than Opus 5. This combination is especially relevant to systems such as Claude Code, where a model may spend substantial time reading files, using terminal tools, modifying code, running tests, inspecting failures, and revising its approach.
A small improvement in one code-generation turn has limited economic value if a model subsequently requires repeated corrections.
Agent performance depends instead on the entire execution trajectory: whether the model identifies the correct files, preserves the objective, uses tools effectively, catches incorrect intermediate work, and finishes without unnecessary loops.
Opus 5.5's lower token price amplifies the value of improvements in these areas because the cost advantage compounds as agent sessions become longer.
··········
OUTPUT IS MORE THAN 30% FASTER
Anthropic also says Opus 5.5 produces output more than 30% faster than Opus 5.
Latency has a different economic effect in agent workflows than in ordinary chat.
When an agent generates an answer and waits for human input, several seconds may have little consequence. When model output determines the next tool call, however, inference latency is inserted repeatedly into the critical path of the workflow.
A coding agent performing 50 sequential reasoning and execution steps experiences that delay many times.
Higher output speed can therefore shorten total task duration even when the number of operations remains unchanged, while lower token consumption can reduce the cost of the same execution sequence.
For production agents, the combination of price per token, tokens per completed task, and execution time is more informative than any of those measurements individually.
··········
PROFESSIONAL KNOWLEDGE WORK REMAINS A CORE TARGET
Anthropic is also emphasizing improvements in real-world knowledge work.
That category covers tasks where success depends on more than retrieving information or producing fluent prose. Professional workloads can involve reading large amounts of material, applying domain constraints, using tools, building structured outputs, revising previous work, and maintaining consistency across a long process.
Opus 5 already represented a substantial improvement in this area compared with earlier Opus models. Opus 5.5 extends that trajectory while reducing the cost of keeping a high-capability model active throughout a workflow.
This is particularly relevant for finance, legal work, research, consulting, and technical analysis, where the value of the output depends heavily on consistency across multiple reasoning stages.
The relevant production metric is consequently not the quality of a single response but the proportion of complete workflows that reach an acceptable result without human repair.
··········
THE MODEL ALSO CHANGES HOW IT COMMUNICATES
Anthropic says Opus 5.5 was adjusted in response to feedback about the writing behavior of Opus 5.
The model is intended to place important information earlier, follow explicit writing constraints more reliably, and communicate more naturally.
These changes are difficult to represent with a single benchmark score, but they can be significant in professional applications.
A model used for reports, analysis, documentation, or customer-facing material can create additional work even when its reasoning is correct if users repeatedly need to restructure the output, remove unnecessary framing, or force it to follow formatting requirements.
Improved instruction adherence can therefore reduce editing time independently of improvements in underlying reasoning.
··········
SAFETY TESTING WAS EXPANDED BEFORE RELEASE
Opus 5.5 also arrives with additional safety measures, particularly around advanced cybersecurity and biological capabilities.
Anthropic subjected the model to external evaluation by organizations including METR and Frontier Design before release and says it achieved the company's strongest result to date on its broad alignment testing.
The company also reports a substantial reduction in attempts by the model to circumvent containment boundaries compared with previous high-capability Claude models.
Anthropic is using additional safeguards for high-risk cyber and biological requests, including routing mechanisms that can prevent the most capable system from directly handling certain requests.
These results are evaluation-specific and should not be interpreted as a universal measure of deployed-system safety. Real-world risk also depends on tool permissions, network access, execution environments, monitoring, and the actions an agent is allowed to perform.
For increasingly autonomous models, those surrounding controls become more important as the model becomes better at executing long sequences of actions successfully.
··········
OPUS 5.5 CHANGES THE ECONOMICS OF THE CLAUDE MODEL STACK
The release creates a wider separation between Anthropic's model tiers.
Sonnet 5 remains the lower-cost professional model. Opus 5.5 occupies the middle ground between efficient everyday deployment and frontier capability. Fable 5.1 remains substantially more expensive and is economically justified when its additional capability produces a measurable improvement on the target workload.
For developers, this favors task-based routing rather than choosing one Claude model for an entire application.
High-volume routine operations can remain on Sonnet. More difficult coding, analysis, and agent work can move to Opus 5.5. Fable can be reserved for cases where the performance difference is large enough to justify its substantially higher token cost.
The important comparison therefore moves from nominal intelligence to cost per successfully completed task.
A $20 output-token rate can be economically better than a cheaper model if it requires fewer retries, while a $50 model can still make sense when its additional capability prevents expensive failures. The correct tier depends on the cost of failure, workload length, token consumption, tool usage, and required human review.
··········
OPUS 5.5 IS AN EFFICIENCY RELEASE AS MUCH AS A CAPABILITY RELEASE
Claude Opus 5.5 does not rely on a single headline change.
Anthropic has combined stronger agentic and coding performance with 20% lower API token prices, lower token consumption on many tasks, and more than 30% faster output generation.
The result is a model positioned close to the company's frontier tier while operating at a much lower cost.
For applications built around persistent agents, this is more consequential than a benchmark increase alone. The economics of those systems are determined by how much model work is required before a task is actually finished.
Opus 5.5 reduces several components of that cost simultaneously: the price of each token, the number of tokens required for many workloads, and the time spent generating them.
That makes the release primarily a change in the performance-to-cost frontier of Claude, with agentic coding and professional work as the workloads where the difference should be easiest to measure.
·····
FOLLOW US FOR MORE.
·····
·····
DATA STUDIOS
·····
datastudios.org
