Anthropic released Claude Opus 4.6 on February 5, 2026, its first Opus-class model with a 1 million token context window in beta and a new Agent Teams capability that lets multiple agents divide and coordinate a single task.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source On coding and agentic benchmarks the model posted 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, at standard pricing of $5 per million input tokens and $25 per million output tokens.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source The stake is the direction of travel: the release is engineered around long-horizon, tool-using work rather than conversation. We assess with moderate confidence that Opus 4.6 marks a real inflection from chat toward autonomous knowledge work, while noting that the headline context number oversells how much of that window models reliably use.
What the release is actually built for
Read the feature list as a statement of intent. The 1 million token beta lets the model hold roughly the equivalent of many hundreds of pages in a single task, a fivefold jump from the prior 200,000 token window, and Anthropic frames it around processing whole codebases and long documents in one pass rather than longer chats.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source Agent Teams extends that from memory to labor: multiple agents work across many tools in one task and recover from errors without a human relaying results between them.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source Adaptive effort controls with low, medium, high, and max settings, plus context compaction that summarizes a conversation automatically as it grows, are the plumbing of long-running jobs, not chat niceties.3 TechInformed 2026-02-10 Opus 4.6 added adaptive effort levels (low, medium, high, max) and context compaction, and was released the same week as OpenAI's GPT-5.3-Codex rollout, with premium pricing above 200,000 tokens. Open source The design target is a model that starts a task, keeps state across it, calls tools, checks itself, and finishes.
The benchmark picture supports the framing without overselling it. Opus 4.6 led agentic coding evaluations with the highest Terminal-Bench 2.0 score recorded at the time and held 80.8% on SWE-bench Verified, while its edge on professional-task measures ran to roughly 144 Elo points ahead of a leading OpenAI model on one knowledge-work benchmark.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source Where it did not lead is as instructive: a competing frontier model kept an advantage on graduate-level scientific reasoning, and Opus 4.6 is priced well above the cheapest frontier options.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source The gains are concentrated where the release points, in agentic and coding work, not evenly across all reasoning.
Second order effects and the ledger
The first order fact is a stronger coding model. The second order effect is pressure on how software work is priced and staffed. If a model can hold a whole repository in context and run a team of agents against a multi-step task, the buyer's question moves from cost per query to cost per completed job, which is where Anthropic wants the comparison to sit given its premium pricing.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source
Who gains: Anthropic, which reinforced a lead in the enterprise developer segment it has made its center of gravity, and which timed the release the same week as a rival's coding rollout to blunt it.3 TechInformed 2026-02-10 Opus 4.6 added adaptive effort levels (low, medium, high, max) and context compaction, and was released the same week as OpenAI's GPT-5.3-Codex rollout, with premium pricing above 200,000 tokens. Open source Enterprises with large codebases and document stores gain the most concrete benefit, because whole-codebase context and self-recovering agents attack exactly the tasks that previously broke context limits.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source Who loses: pure chat-first assistants and thinner application-layer wrappers, whose differentiation erodes as the base model absorbs multi-agent orchestration and long context that startups previously stitched together. Cost-sensitive buyers also lose relative ground, since the capability arrives at $5 and $25 per million tokens rather than at the bottom of the market.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source We assess with moderate confidence that the biggest displaced value is in tooling built to paper over short context windows, which the model now does natively.
The counter-case
The strongest argument against the autonomy narrative is that a large context window is not the same as reliable use of it. Models routinely degrade in the middle of very long inputs, and a 1 million token beta buys headroom, not guaranteed recall across the whole span.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source Agent Teams likewise multiplies both capability and failure surface: more agents and more tool calls mean more places for an error to compound before a human sees it, and self-recovery claims are exactly the kind that hold in demos and fray in production.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source The thesis that this is an inflection toward autonomous work fails if, in practice, teams keep humans tightly in the loop because the long-context and multi-agent features prove brittle on real tasks, leaving Opus 4.6 as a better assistant rather than an autonomous worker. That the model still trailed a rival on hard scientific reasoning is a reminder that the frontier moved unevenly, not uniformly.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source
What to watch
- The 1M window leaves beta. If Anthropic promotes the 1 million token context to general availability with published retention benchmarks by mid-2026, the long-context claim is real; a prolonged beta with caveats would signal the window is headroom more than dependable memory.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source
- Agent Teams shows up in named deployments. Concrete enterprise case studies of multi-agent tasks running unsupervised within two quarters would validate the autonomy thesis; their absence would suggest teams stayed human-supervised.1 SiliconANGLE 2026-02-05 Claude Opus 4.6 released February 5, 2026 with a 1 million token context window in beta (up from 200,000) and Agent Teams that divide work across tools and recover from errors. Open source
- Pricing holds as rivals cut. If Anthropic sustains $5 and $25 per million tokens while cheaper models close the coding gap, it is betting completed-task value over price; a cut would concede the comparison moved to cost per query.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source
- The coding lead survives the next OpenAI and Google releases. Whether Opus 4.6's Terminal-Bench and SWE-bench edge holds through the next frontier launches will show if agentic coding is a durable Anthropic moat or a temporary cadence advantage.2 Fello AI 2026-02-06 Opus 4.6 scored 65.4% on Terminal-Bench 2.0 and 80.8% on SWE-bench Verified, priced at $5/$25 per million tokens, leading agentic work but trailing a rival on graduate-level reasoning and costing more than the cheapest models. Open source 3 TechInformed 2026-02-10 Opus 4.6 added adaptive effort levels (low, medium, high, max) and context compaction, and was released the same week as OpenAI's GPT-5.3-Codex rollout, with premium pricing above 200,000 tokens. Open source The signal to track is not the launch benchmark; it is whether autonomous multi-step jobs start replacing supervised ones inside real engineering teams.