Morning Edition · Wednesday, July 8, 2026
Anthropic Ships Claude Sonnet 5 With a One-Million-Token Context Window
The company reports 82.1 percent on SWE-bench Verified, though independent trackers list a range of scores depending on the coding scaffold used.
Anthropic released Claude Sonnet 5 on June 30, positioning it as its most agentic mid-tier model and extending its context window to one million tokens, five times the prior 200,000-token limit. It is now the default for Free and Pro users and is available in Claude Code, the API, Amazon Bedrock, Google Vertex AI, and third-party coding tools.
Anthropic reported 82.1 percent on SWE-bench Verified, which the company frames as the first model to cross 80 percent at launch. That figure warrants caution. Independent benchmark aggregators list Sonnet 5 across a wider range, with Vellum recording 72.7 percent on the same test and placing the model below Opus 4.8's 79.4 percent. The gap is not fraud, it is scaffold sensitivity. SWE-bench scores move several points depending on the agent harness, retry budget, and test-time compute, which is exactly why a single vendor number should not be treated as settled.
More telling is the pattern of the release. Sonnet 5 reportedly scores higher than the larger Opus 4.8 on Terminal-Bench 2.1, a coding benchmark, while trailing on OSWorld and SWE-bench Pro. Introductory pricing is $2 per million input tokens and $10 per million output through August 31, then $3 and $15.
The strategic point is that Anthropic is compressing near-frontier coding capability into a cheaper tier, the same strategy that keeps coding central to competition among the labs.
What this means
The exposed parties are rival labs selling coding capability at premium tiers. By pushing an 80-plus SWE-bench claim into a $2 input model with a million-token window, Anthropic pressures the price-per-capability of OpenAI's Codex and Google's Gemini coding stack through the cost channel, not just raw capability. The honest uncertainty is the benchmark. If independent harnesses converge near the low-70s rather than 82 percent, this is a solid iterative release, not a threshold break. What decides it is reproducible third-party runs on a fixed scaffold.
What to watch
- Independent SWE-bench Verified reproductions on a standardized harness, which would confirm or undercut the 82.1 percent vendor figure.
- Whether the one-million-token window holds effective retrieval accuracy across the full context or degrades in the middle, the real test of long-context coding utility.
Observations to monitor, not financial advice.
Synthesized from: Anthropic News · Vellum
Part of a tracked trend
Frontier Labs Race on AI Coding Capability
Coding is becoming a primary competitive battleground among frontier labs, with incumbents standing up permanent coding teams and investing in new training stages (e.g. midtraining) to match leaders like Anthropic; expect recurring reorganizations, benchmarks, and model releases aimed specifically at code.
More from this edition
- China Weighs Curbing Overseas Access to Its Most Advanced AI Models
- Meta's Non-Invasive Brain-to-Text Decoder Reaches 61 Percent Word Accuracy
- NVIDIA Launches Vera, a CPU Built for the Agent Loop
- Cloudflare Opens a Stablecoin Payment Rail for AI Agents at the Network Edge
- NVIDIA and Hugging Face Push Isaac GR00T 1.7 Into the Open LeRobot Stack
- A 1B Vision Model Beats 7B Rivals on Dense Spatial Perception
- Google Expands Managed Agents in the Gemini API With Background Tasks and Remote MCP
- Anthropic Redeploys Fable 5 and Proposes a Cross-Lab Jailbreak Severity Scale
- Report of a Cross-Tenant Data Leak in Claude Code Tests a Core Isolation Guarantee
- New Benchmark Tackles the Chaos in Comparing Long-Context KV-Cache Optimizations
- Cloud and AI Vendors Flood Startups With Free Compute Credits to Lock Them In