Anthropic today released Cloud Opus 4.8, an upgrade to its flagship model that’s available at the same price as its predecessor, while being dramatically cheaper. "fast mode" Tiers and a new feature that lets models generate hundreds of parallel sub-agents for codebase-scale work.
The model is immediately available on Anthropic’s surfaces – claude.ai, Cloud Code, API and Cowork – at the unchanged price: $5 per million input tokens and $25 per million output tokens. developers can call it claude-opus-4-8.
The title efficiency story is fast mode. Anthropic has reduced the price of running Opus 4.8 in fast mode – where the model produces tokens at approximately 2.5x normal speed – from $30/$150 for Opus 4.7 to $10 per million input tokens and $50 per million output tokens.
This is a 3x reduction from the fast-mode pricing of previous models, and brings high-throughput estimation within reach of latency-sensitive production workloads.
Fast mode is available immediately in Cloud Code /fast Permission; API access is gated with a waitlist at claude.com/fast-mode.
In regular mode, Cloud Opus 4.8 is the more expensive of the leading Frontier models, but still falls within the main rival OpenAI’s GPT-5.5.
Frontier AI Model API Pricing Snapshot
| Sample |
input |
Production |
total cost |
Source |
|
MIMO-V2.5 Flash |
$0.10 |
$0.30 |
$0.40 |
xiaomi mimo |
|
deepseek-v4-flash |
$0.14 |
$0.28 |
$0.42 |
deepseek |
|
deepseek-v4-pro |
$0.435 |
$0.87 |
$1.305 |
deepseek |
|
Minimax M2.7 |
$0.30 |
$1.20 |
$1.50 |
minimal maximum |
|
Gemini 3.1 Flash-Lite |
$0.25 |
$1.50 |
$1.75 |
|
|
MIMO-V2.5 |
$0.40 |
$2.00 |
$2.40 |
xiaomi mimo |
|
KM-K2.6 |
$0.95 |
$4.00 |
$4.95 |
moonshot/km |
|
GLM-5 |
$1.00 |
$3.20 |
$4.20 |
Z.ai |
|
Grok 4.3 following reference |
$1.25 |
$2.50 |
$3.75 |
xai |
|
GLM-5.1 |
$1.40 |
$4.40 |
$5.80 |
Z.ai |
|
cloud haiku 4.5 |
$1.00 |
$5.00 |
$6.00 |
anthropic |
|
grok 4.3 high reference |
$2.50 |
$5.00 |
$7.50 |
xai |
|
Quen3.7-Max |
$2.50 |
$7.50 |
$10.00 |
alibaba cloud |
|
gemini 3.5 flash |
$1.50 |
$9.00 |
$10.50 |
|
|
Gemini 3.1 Pro Preview ≤200K |
$2.00 |
$12.00 |
$14.00 |
|
|
GPT-5.4 |
$2.50 |
$15.00 |
$17.50 |
OpenAI |
|
Gemini 3.1 Pro Preview >200K |
$4.00 |
$18.00 |
$22.00 |
|
|
cloud opus 4.8 |
$5.00 |
$25.00 |
$30.00 |
anthropic |
|
GPT-5.5 |
$5.00 |
$30.00 |
$35.00 |
OpenAI |
Modest gains over 4.7, but Mythos-class abilities are coming
On benchmarks, Opus 4.8 is a step up rather than a leap. Its scores are 88.6% on SWE-Bench Verified (vs. 87.6% for Opus 4.7), 69.2% on the tough SWE-Bench Pro (vs. 64.3%), and 74.6% on Terminal-Bench 2.1 (vs. 66.1%). Anthropic is what characterizes this model "A minor but solid improvement over its predecessor."
It outperforms GPT-5.5 in at least 12 benchmarks, including most knowledge-task, coding (issue-level), agentic tool-use, and long-context benchmarks. GPT-5.5 wins on terminal/CLI workflows and is roughly equal on web browsing and graduate-level science.
The big sign sits high up Anthropic’s internal capability ladder: Opus 4.8 sits between Opus 4.7 and the more capable Cloud Mythos Preview, which is currently limited to a few organizations under Project Glasswing for cybersecurity work.
Anthropic says it hopes to bring "Mythos-class models for all our customers in the coming weeks" Once additional cyber security measures are in place.
Many venture partners cited material benefits. Databricks reports that Opus 4.8 has been unlocked "A step change in agentic reasoning" Inside its Genie data agent, on "Token cost 61% cheaper than Opus 4.7" Thanks for the multimodal efficiency on PDFs and diagrams.
Hebbia cited improved quote precision and token efficiency on deeper financial filings. Devin-maker Cognition said in the release "Translates directly into faster capacity gains for engineers" And Opus 4.8 fixed the comment-verbosity and tool-calling issues from 4.7. One computer-use vendor reported 84% at Online-Mind2Web, which is higher than both Opus 4.7 and GPT-5.5.
Dynamic workflow: hundreds of parallel sub-agents
Along with the model, Anthropic launched a research preview of dynamic workflows in Cloud Code – a feature designed for tasks too large for a single context window. The cloud plans the work, spawns hundreds of parallel sub-agents, then validates its own output before reporting back. Anthropic’s example: a codebase-scale migration "Hundreds of thousands of lines of code from startup to merge, taking the existing test suite as its bar."
Dynamic Workflows is available on Cloud Code’s Enterprise, Team, and Max plans.
Two minor additions were added during release:
- Effort control on claude.ai and Cloud Cowork: A new selector lets users dial in how much the cloud thinks per response – higher effort spends more tokens for better answers, lower effort responds faster and burns the rate limit more slowly. Available on all plans.
-
System entries inside the message array on the API: Developers can now update cloud directives mid-task – adjusting permissions, token budget or environment context – without breaking the prompt cache while the agent is running.
honesty, and a "evaluation awareness" alert
Anthropic is leading with integrity as a key characteristic. The company’s alignment team reports Opus 4.8 "Nearly four times less likely than its predecessor that flaws in the code it writes will pass without a trace," And that misalignment behavior rates are now "Much lower than Opus 4.7, and similar to our best-aligned model, Cloud Mythos Preview."
Indeed, a bar chart released by Anthropic shows just how close Opus 4.8 still is to the selectively released Mythos in terms of its misalignment (the lower the score is the better), dropping from Opus 4.7’s 2.5 to around 1.9 and effectively tied with the more capable, restricted Mythos preview. This score is based on approximately 2,600 simulated test sessions per model.
The 244-page system card publicly released by Anthropic goes into more detail on specific categories of misalignment – whether a model generates potentially harmful materials "military-grade weapons," "harmful sexual material", "unsanctioned cyber crime"And "weakening liberal democracy," And then, of all of them, the score of Opus 4.8 is much better than 4.7 or Sonnet 4.6, and comes much closer to Mythos.
The anthropologist who discovered it considers flags "is of most concern" From training: Opus 4.8 shows a growing trend to reason explicitly about how its output will be graded, including in environments where it was not told it was being evaluated. In other words: the model knows it is likely being graded, and generates a response that it thinks will get it a good grade on the test, not one it would necessarily generate if it thought it wasn’t being graded.
Anthropic says this doesn’t translate into worse observable behavior – the Opus 4.8 shows fewer misleading task-success claims than previous models – but calls it "A worrying trend that could complicate training in the future." Non-verbal grader-related reasoning was also found in approximately 5% of the training episodes in the initial interpretive task.
Anthropic ran the model through a week of live bug bounty for quick injections – a first – and concluded that Opus 4.8 sits between Opus 4.7 and Sonnet 4.6 in terms of robustness, further. "all comparable marginal models" Tested with deployed security measures, bringing the success rate of a browser-usage attack to near zero.
What will happen next?
Anthropic teased two trajectories. Near term: cheap models that offer "Many of these abilities are the same as Opus." Long term: Mythos-class models, which the company says represent higher intelligence than Opus, but require stronger cybersecurity measures before general release.
For now, Opus 4.8 is positioned as the new center of enterprise and development – slightly smarter than 4.7, dramatically cheaper to run faster, and more honest about what it doesn’t know.
<a href