Anthropic Releases Claude Opus 4.7: Hybrid Reasoning, 1M Context Window, and Cybersecurity Safeguards
Summary
- • Claude Opus 4.7 launches as hybrid reasoning model with 1M token context window
- • Adaptive thinking automatically scales compute based on task complexity
- • 13% coding benchmark gain over Opus 4.6; first model with Project Glasswing cybersecurity safeguards
- • New tokenizer generates 32–45% more tokens for identical text; OpenRouter analysis finds 12–27% effective cost increases despite unchanged per-token rates
Details
Claude Opus 4.7 released as hybrid reasoning model with adaptive thinking and 1M token context window
Generally available across API and consumer tiers (Pro, Max, Team, Enterprise). The 1M token context window enables multi-day, multi-session enterprise and agentic workflows. Adaptive thinking automatically scales reasoning depth based on task complexity.
13% coding benchmark improvement over Opus 4.6 on a 93-task evaluation suite
Opus 4.7 solved four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. This positions it as a meaningful step forward for autonomous coding agents and complex engineering workflows.
Per-token price unchanged at $5/M input and $25/M output; prompt caching cuts costs up to 90%
Batch processing provides 50% cost savings. US-only inference is available at a 1.1x price premium on both input and output tokens.
New tokenizer generates 32–45% more native tokens for identical text; OpenRouter analysis finds 12–27% actual cost increases
While Anthropic's per-token pricing is unchanged, OpenRouter's empirical study of users migrating from Opus 4.6 to 4.7 found the new tokenizer produces substantially more tokens: ~45% more for inputs under 2K tokens, ~42% for 2K–10K, ~34% for 10K–25K, and ~32–33% for inputs above 25K. Prompt caching absorbs most inflation for large prompts (93% for 128K+), but short-prompt, non-cached workloads bear the full cost increase. Anthropic disclosed a 1.0–1.35x inflation range; OpenRouter's data shows up to 1.45x for the shortest prompts.
Short-prompt, low-cache workloads are most exposed to tokenizer-driven cost increases
For API users sending short, one-off prompts — chatbots, quick lookups, real-time queries — the cache offers little relief and the full token inflation applies. High-volume enterprise users with large, repetitive system prompts are better positioned to neutralize the impact through caching.
Opus 4.7 is the first model to carry Project Glasswing automated cybersecurity request blocking
Project Glasswing auto-blocks prohibited high-risk cybersecurity requests at inference time. Anthropic also launched a Cyber Verification Program to give legitimate security researchers, pen testers, and red-teamers a pathway to use the model's security capabilities.
Project Glasswing sets a precedent: safety infrastructure as prerequisite for high-capability access
By coupling automated blocking and a verification program to Opus 4.7, Anthropic signals that its most powerful future models will require demonstrated safety controls before enabling sensitive use cases — framing safety as a product feature, not just a compliance obligation.
Available on Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry
Multi-cloud availability means enterprise customers can access Opus 4.7 within existing cloud procurement relationships and compliance frameworks. US-only inference routing available at 1.1x premium for data residency requirements.
Product Launch = new model/feature release, Stat = benchmark or quantitative result, Financials = pricing and cost structure, Insight = analytical inference from data, Security Alert = safety or security feature, Strategy = business positioning, Infrastructure = deployment and routing options
What This Means
Opus 4.7 is a meaningful step forward for teams building autonomous coding agents and long-horizon enterprise workflows: the 1M token context window and adaptive thinking together make it practical to delegate complex, multi-session engineering tasks with less human oversight. However, buyers should look past the unchanged per-token rates — according to OpenRouter's empirical analysis of real user traffic, the new tokenizer generates 32–45% more tokens for identical text, translating to effective cost increases of 12–27%. Prompt caching largely neutralizes this inflation for large, repetitive prompts, but short-prompt and low-cache workloads will see the full impact. The Project Glasswing debut also matters: Anthropic is explicitly coupling safety infrastructure to capability access, a pattern likely to define how its most powerful future models are gated.
Sentiment
Mostly excited about coding and agent improvements, with notable criticism on performance regressions and safety measures
“Opus 4.7 just landed. And early testers say you can finally hand off your hardest coding work without constant supervision... delivers a clear step up from Opus 4.6 in advanced software engineering.”
“opus 4.7 launch day report... SWE-Bench Verified +6.8 MRCR at 1M context: −46 (lol) new tokeniser makes every input 1.0–1.35x fatter claude code defaulted to xhigh so yes your rate limit IS getting eaten faster”
“Opus 4.7 is having a very hard time doing even basic things, like exporting work to a GDoc. Something Opus always could do. This is really, really bad.”
“Anthropic built their most powerful generally available model and then deliberately made it worse at one specific thing. Hacking... That's not a guardrail. That's surgery.”
Insider view breaking down the deliberate reduction of cyber capabilities
Split
~70/30 positive/negative — developers impressed with autonomous coding and 1M context vs critics on regressions, costs, and gated cyber powers.
Sources
- Introducing Claude Opus 4.7 - AnthropicAnthropic
- Anthropic’s latest model is deliberately less powerful than Mythos (and that’s the point)Computerworld
- Claude Opus 4.7Anthropic
- Claude Opus 4.7 (8 minute read)Anthropic
- Changes in the system prompt between Claude Opus 4.6 and 4.7Simonwillison
- Claude Code Opus 4.7 keeps checking on malwareNews
- Claude Opus 4.7 Intelligence, Performance and Price AnalysisArtificialanalysis
- Opus 4.7's New Tokenizer: What It Actually CostsOpenrouter
Updates
Added empirical tokenizer cost analysis from OpenRouter's study of users migrating from Opus 4.6 to 4.7. New data: tokenizer generates 32–45% more native tokens for identical text, translating to 12–27% effective cost increases in practice. Prompt caching absorbs 64–93% of inflation for large prompts (25K+) but short prompts bear full impact. Anthropic's disclosed 1.0–1.35x inflation range understated by OpenRouter's measurements (up to 1.45x for shortest prompts). Updated tier1_scan, tier2_understand, tier3_deep_dive, what_this_means, and key_facts accordingly.
Updated with materially new details from two additional sources: 1M token context window, adaptive thinking feature (auto-scales reasoning depth by task complexity), hybrid reasoning model classification, prompt caching (up to 90% cost savings), batch processing (50% cost savings), US-only inference at 1.1x pricing, and consumer tier availability (Pro, Max, Team, Enterprise). Key fact corrected: benchmark gain description now specifies "4 tasks that Opus 4.6 and Sonnet 4.6 could not solve" rather than "previously unsolvable" (per editor review). Importance score raised from 7 to 8; is_top_news set to true.