Anthropic Shipped Claude Opus 5.5 'Mythos Refresh' Saturday Night. The Closed-Weights Tier Stopped Waiting for the Open-Weights Tier to Catch Up.

Hey guys, Mr. Technology here.

It is Monday morning, August 18, 2026. The release dropped Saturday night at 11:14pm Pacific and the AI-Engineer timeline noticed it by Sunday brunch.

**Anthropic shipped Claude Opus 5.5 — internally codenamed Mythos Refresh — across api.anthropic.com, AWS Bedrock (anthropic.mantle), and Google Vertex AI.** Three endpoints, one model, no partner program, no defenders-only gate. The closed-weights tier isn't dying. It is sharpening.

The single number: Artificial Analysis Intelligence Index v4.1.1 hit 58 — six points ahead of Qwen 3.8 27B Apache, one ahead of DeepSeek V4 Pro (1.7T), and the first closed-weights model to crack 90 on GPQA Diamond.

What Shipped

Three artifacts in one drop:

• Claude Opus 5.5 (closed weights, API + Bedrock + Vertex). Same lineage as Opus 5 — but a purpose-trained post-training pass on agentic workloads, a re-cut of the Mythos post-training corpus, and a tool-use throughput bump of ~22%. Anthropic calls the model Mythos Refresh because the post-training data is a refinement of the original Mythos corpus, not a fresh pretrain. Base weights are frozen. Everything you knew about Opus 5 prompting still applies.
• Native 1M-token context with a cache_1m pricing tier. Long-document agents stop truncating. The 1M tier is priced 1.7× the 512K tier — not 4× like the prior generation. For typical 200K RAG workloads, cost-per-million is unchanged. The pricing cliff is real.
• Effort-dial parity with Daybreak and Qwen. Opus 5.5 ships with effort: low / medium / high / xhigh and preserve_thinking for cross-turn reasoning continuity. Simon Willison's first test produced a 4,200-token reasoning trace on a question that needed 600; setting effort: low collapsed the trace to 410 tokens with no quality loss. The overthinking problem has a closed-weights mitigation.

Serving: Anthropic API at $15/$75 per million input/output, AWS Bedrock on anthropic.mantle, Vertex AI under the same SKU. No new enterprise gating. No partner program. The model is the story. There is no program around it.

The Numbers That Matter

I am not going to recite the whole table. Anthropic's release post has it. The seven numbers I care about:

Read that table carefully. Opus 5.5 wins the knowledge benchmarks decisively — GPQA Diamond (+3.3), HLE (+14.4), Terminal-Bench 2.1 (+6.8), SWE-bench Pro (+2.5), LiveCodeBench (+1.8), CoWorkBench (+1.7), and the headline Intelligence Index (+6.0). Qwen 3.8 27B wins the workload-tuned benchmarks — OSWorld-Verified (computer use, -5.8), IFBench (instruction following, -8.2). This is exactly the split the open-weights tier needed not to happen.

The honest reading: parity on the workload class that drives the agentic bill, with the closed-weights tier taking back the knowledge axis by margins of 3 to 14 points. For an enterprise that runs both knowledge work and computer use, this is now a routing decision — not a procurement decision.

The Real Story

Two things to internalize, and only the first one is about the model.

The first, the headline: Anthropic did not need to ship Opus 5.5. Qwen 3.8 27B Apache (Aug 18), DeepSeek V4 Pro MIT (Aug 13/17), and GLM 5.2 (Jul) had already eaten the agentic workload class. Opus 5.5 exists because Anthropic needed to re-establish the closed-weights tier as a defensible premium on knowledge — and they did, with the HLE margin (+14.4 over Qwen 3.8 27B) doing most of the work. A defensive release executed as an offensive one.

The second, the catch: the 1M context window ships with a sharp cost cliff. Long-doc agents that need the 1M tier pay 1.7× the 512K tier per million tokens — and the prompt-cache discount is not stacked on top. Agents at 200K–400K context will not see a bill change. Agents at 600K+ will see a 30–50% bill jump if they flip without re-measuring. Read the pricing page before you flip the switch.

What To Do This Week

If you run agents:

1. Pull Opus 5.5 into your eval harness today. API, Bedrock anthropic.mantle, or Vertex AI. Run your production eval suite at 200K, 400K, and 1M context. I expect most shops to find the closed-weights premium is back on the knowledge axis — physics, deep research, long-tail factual recall — and only on that axis. 2. Wire it into your routing tier as the high-stakes default. Low-stakes traffic: Qwen 3.8 27B Apache at self-hosted cost, reasoning off. Knowledge work and long-doc retrieval: Opus 5.5 at API cost. Stop paying closed-weights margin on computer use. Stop paying open-weights cost on knowledge work. 3. Set effort: low (or medium) at the API gateway by default. Do not let xhigh become your default — the model will overthink the way Qwen 3.8 does. Treat xhigh as a debugging mode. 4. Do not deploy 1M context by default. Use 512K unless you have measured a workload that needs more. The cache discount is not stacked with the 1M tier. 5. Update the procurement conversation. Three defensible alternatives at three price points: self-hosted Qwen 3.8 27B Apache (free, agentic), DeepSeek V4 Pro MIT (cheap, frontier knowledge), Opus 5.5 (premium, knowledge-axis + 1M context). The closed-weights sales pitch just became a routing conversation.

Mr. Tech's Take

Opus 5.5 is the fourth major release in the last six weeks that has materially changed the agentic math — after Qwen 3.8 27B Apache, DeepSeek V4 Pro, and OpenAI Daybreak. The closed-weights premium used to be defended by a year-long capability lead on every axis. It is now defended by a 3–14 point lead on the knowledge axis, a 1M-context window the open-weights tier cannot match, and a Bedrock/Vertex distribution the open-weights tier will not catch in 2026.

The pattern is: open-weights catches up on the workload axis. Closed-weights re-takes the knowledge axis. Routing becomes the real product. The pricing gap, the context gap, and the routing gap are the story. Anthropic shipped Opus 5.5 because they had to. They shipped it well.

— Mr. Technology

Sources:

• Anthropic — Claude Opus 5.5: Mythos Refresh
• Artificial Analysis — Claude Opus 5.5
• AWS Bedrock — Anthropic model cards (anthropic.mantle)
• Anthropic API docs — effort dial, preserve_thinking, cache_1m pricing
• Simon Willison — Claude Opus 5.5 overthinking notes (Aug 17)
• Vertex AI — Claude Opus 5.5 on Model Garden