1M Context Window Goes GA for Opus 4.6 and Sonnet 4.6 — No Premium, 6x More Media
Anthropic made the full 1M token context window generally available for Claude Opus 4.6 and Sonnet 4.6 at standard API pricing — no long-context premium, no beta header. A 900K-token request costs the same per-token rate as a 9K one: $5/$25 per million tokens for Opus 4.6, $3/$15 for Sonnet 4.6. Requests over 200K tokens that previously required a beta header now work automatically.
The media limit also jumped from 100 to 600 images or PDF pages per request when using the full context window — a 6x increase that matters for document-heavy enterprise workflows. Available today on Claude Platform, Microsoft Azure Foundry, and Google Cloud Vertex AI. For Claude Code users, 1M context is now included on Max, Team, and Enterprise plans with Opus 4.6. Anthropic reports Opus 4.6 scores 78.3% on MRCR v2 at 1M tokens, showing the model holds up across the full window.
Claude Certified Architect: Foundations — Anthropic’s First Technical Credential
Alongside the Partner Network launch, Anthropic quietly shipped its first-ever technical certification: Claude Certified Architect, Foundations. It’s a technical exam for solution architects building production applications with Claude. The cert covers architecture patterns, prompt design for production, tool use, and agentic workflow design. Additional certifications are planned for later in 2026.
This matters because it creates a standardized credential that consulting firms and SIs can use to staff Claude projects with confidence. Partners joining the Claude Partner Network get priority access. For individual devs, it’s a legible signal that differentiates Claude expertise from general AI experience on a resume.
Automatic Caching, Fine-Grained Tool Streaming, and Free Code Execution Now Live
A batch of API improvements landed this week. Automatic caching for the Messages API means developers no longer need to manually manage cache breakpoints — the system handles it. Fine-grained tool streaming is now GA on all models and platforms, giving real-time visibility into tool call progress. And sandboxed code execution is free when paired with web search or web fetch, cutting costs for agentic pipelines that combine reasoning with retrieval.
Web search and programmatic tool calling are also fully GA — no beta header required. Dynamic filtering via code execution lets models filter search results before they land in the context window, meaning leaner context and lower token bills on search-heavy tasks. These are incremental improvements individually, but together they meaningfully reduce the friction and cost of building production Claude agents.
Claude Haiku 3 Retires April 19 — 36 Days to Migrate
A reminder that the clock is ticking on claude-3-haiku-20240307. Anthropic confirmed the hard retirement date of April 19, 2026. Any app or pipeline still routing to this model string will break on that date. The migration target is Claude Haiku 4.5, which is available now. If you’re running high-volume classification, routing, or summarization at scale, test the swap sooner rather than later — Anthropic has not historically extended these deadlines.
Palantir Defies Pentagon Blacklist, Keeps Using Claude
Despite the Pentagon’s supply chain risk designation, defense contractor Palantir continues to use Anthropic’s Claude models in its products, according to reporting from Invezz. This is significant because it highlights the gap between the Pentagon’s policy and the reality on the ground: Claude is so deeply embedded in military and intelligence workflows that even a formal blacklist hasn’t forced immediate removal. Palantir’s continued use also puts the company in an awkward position — technically defying a Pentagon directive while serving as one of the DoD’s most important technology partners.
The situation underscores what Fortune reported last week: senior Pentagon officials had a realization about how indispensable Anthropic had become to defense operations. Claude is currently the only AI model deployed on classified Pentagon systems, making a clean removal far more complex than the 180-day memo suggests.
Anthropic Revenue on Pace to Surpass OpenAI by End of 2026
Anthropic’s financial trajectory is accelerating. The coding agent alone was running at $1B annualized by end of 2025; by February 2026 that had more than doubled to $2.5B. Combined with the consumer surge (11M daily users, #1 app in 15 countries), enterprise self-serve plans, and the new Marketplace, Anthropic is now on track to surpass OpenAI’s total revenue by end of year. That’s a stunning reversal from twelve months ago, when OpenAI had a roughly 10x revenue lead.
Pentagon Memo Orders 180-Day Removal of Claude from All Military Systems
An internal Pentagon memorandum dated March 6 has formally ordered senior military commanders to remove Anthropic’s AI products from their systems within 180 days, CBS News and The Hill confirmed. The memo follows the unprecedented supply chain risk designation and applies across all branches. Defense Secretary Hegseth’s office is enforcing the directive, and defense contractors are also being notified they must certify their systems are Claude-free.
The practical challenge is enormous. Claude is currently the only AI model deployed on the Pentagon’s classified systems, and multiple commands are actively using it in the Iran theater. A Syracuse University analysis published March 13 argues the dispute reveals the fundamental limits of AI self-regulation — Anthropic drew ethical lines, but the government is now punishing it for exactly that. The 180-day timeline puts the removal deadline in September, well after Anthropic’s lawsuit is expected to see its first substantive court hearing. If Anthropic wins an injunction, the memo may never be enforced.
The 1M Context Play Is a Pricing Landgrab
Making 1M context GA at standard pricing is a competitive move disguised as a product update. Most frontier models either don’t support context windows this large, charge a premium for them, or degrade meaningfully at the top end. Anthropic is betting that commoditizing long context will pull enterprise workloads that currently rely on RAG pipelines, external vector stores, or chunking strategies — replacing complexity with brute-force context. At 78.3% on MRCR v2 at the full million tokens, the model actually holds up.
The 6x media increase (600 images/PDFs per request) is the enterprise unlock. Legal, financial, and compliance teams that need to process entire document sets in one pass now have a viable path. Pair that with automatic caching and free code execution on search queries, and the cost of building production agents on Claude just dropped materially. Anthropic is making it economically irrational to not use the full context window — which also makes it harder to switch to competitors who still charge more for less.