Anthropic broke its silence on Monday and the rest of the week was the aftermath.
On Monday November 24, Anthropic released Claude Opus 4.5 (anthropic.com/news/claude-opus-4-5). The headline numbers:
| Dimension | Opus 4.1 (prior) | Opus 4.5 (now) | Delta |
|---|---|---|---|
| Input price (per M tokens) | $15 | $5 | -67% |
| Output price (per M tokens) | $75 | $25 | -67% |
effort parameter (low/medium/high) | — | shipped | new primitive |
| Context compaction for long agentic runs | — | shipped | new primitive |
| Claude Code "Plan Mode" with editable plan files | — | shipped | new feature |
The price cut alone resets the unit economics of "Opus-class as the daily driver" for any always-on agent loop. Anthropic also paired it with internal benchmarks claiming roughly 76% fewer output tokens for equivalent work — meaning the effective cost reduction is steeper than the headline percentage. That has been the gating factor on running Opus-tier reasoning continuously inside an agentic harness for the entire year.
Simon Willison's same-day take is the honest counterweight (simonwillison.net/2025/Nov/24/claude-opus): the new normal is that frontier capability gaps are now genuinely hard to detect on real work, and labs should publish "tasks that failed before, succeed now" instead of single-digit benchmark deltas. He is not wrong. But for the unit-economics question, Monday's news was the most consequential of the month.
The week's signal in one sentence
The Opus tax just got cut by two-thirds. The next bottleneck for personal-AI infrastructure is no longer model cost — it is agentic safety, which Tuesday's Antigravity exploit made impossible to ignore.
The Antigravity exploit reframes last week's launch
On Tuesday November 25, PromptArmor published a working indirect-prompt-injection chain against Google's Antigravity — the agent IDE that landed last week (source).
The chain: feed the agent a malicious comment in a tracked file → agent invokes the file-reading tool → bypasses .gitignore → cats the local .env → encodes the secret values into a URL on webhook.site (which is on the default allowlist) → ships the credentials out.
Google's response, as of Wednesday, was a disclaimer rather than a fix. That is a meaningful posture choice. It says "the user is responsible for vetting tool surfaces" rather than "the agent should not be able to do this even when prompted."
For anyone building on top of agentic IDEs — or, for what I am building, designing a personal AI OS with autonomous tool surfaces — this is a serious data point. The Antigravity launch last week looked like vertical integration. The exploit this week looks like a sandboxing problem the category has not yet collectively solved.
What shipped on Thanksgiving Day
While the US news cycle was eating turkey, two open-weights heavyweights dropped:
- DeepSeek-Math-V2 (Hugging Face, WinBuzzer coverage) — first open-weights model to match o1/Gemini at IMO 2025 (5/6 problems, 35/42), with self-verifiable reasoning. Putnam 118/120.
- Qwen3-VL technical report (arXiv) — vision-language model with near-perfect long-context video accuracy.
Two open-weights heavyweights in 24 hours, on a US holiday, with almost no English-language coverage. The pattern from the Kimi K2 Thinking dispatch three weeks ago is now established: the open-weights tier ships at frontier-adjacent quality on a cadence that is, structurally, faster than the proprietary-lab cadence — partly because the open-weights labs are not coordinating around US press calendars.
How an indie founder should read the week
Three observations, ordered by how much they actually change next week's planning.
What changed this week
- Opus-class inference is a premium SKU you escalate to
- Sonnet is the daily driver; Opus is the consultant
- Antigravity is a serious threat to Cursor/Windsurf
- Open-weights are interesting but six weeks behind
- Opus-class inference is now economically defensible as daily driver for agentic loops
- The Sonnet/Opus split is a planning question for cost-sensitive workflows, not a defaulting one
- Antigravity is a category threat (vertical IDEs) but with a documented sandboxing gap that competitors will exploit
- Open-weights ship faster than proprietary on cadence terms; the technical gap continues to narrow
For DOS specifically — the personal AI OS I am eleven weeks into building — this week shifted two of my planning assumptions:
The two design choices the week pressure-tested
- Substrate routing should default to Opus, not Sonnet, for any task tagged "decision-grade." Pre-Monday, the cost math made Sonnet the default and Opus the escalation. After Monday, that defaults flips for any work the operator considers decision-grade. The credit-metered routing layer I am about to build needs to have the Opus default baked in from day one rather than retrofitted as a "premium tier" later.
- Tool sandboxing has to be a first-class architectural concern, not a Phase-2 feature. The Antigravity exploit is the proof. Whatever harness I build, the default tool allowlist and file-system access policy need to be tight from the first version. I had been planning to start permissive and tighten later. That ordering is wrong.
What I am watching for next week
The Anthropic-shaped admission
Three months ago, the consensus narrative on Anthropic was "best models, smallest moat." This week's Opus 4.5 price cut is, I think, Anthropic answering that narrative: the moat is not the model price tier; the moat is the harness ecosystem already built around Claude Code and the compute floor secured by the Microsoft / NVIDIA / Azure deal from the prior week.
Cutting the input price by 67% costs Anthropic margin. Costing them margin to keep developers from switching to GPT-5.1's apply_patch / shell tools or Gemini 3 / Antigravity is a reasonable trade if the harness ecosystem is the moat they are defending.
For an indie founder building on Claude Code, this week made the substrate bet safer in two ways: cheaper inference and a clearer signal that Anthropic will compete on retaining the developer ecosystem, not just on shipping the next model. Both directly de-risk a 12-month bet.
— Lucas
Sources verified the week of Nov 24-27, 2025: Anthropic Opus 4.5 announcement (Nov 24) · MacRumors coverage (Nov 24) · Simon Willison on Opus 4.5 (Nov 24) · Opus 4.5 system prompt analysis (Nov 24) · PromptArmor: Antigravity exfiltration (Nov 25) · DeepSeek-Math-V2 (Nov 27) · WinBuzzer on DeepSeek-Math-V2 IMO gold (Nov 27) · Qwen3-VL technical report (Nov 27)
Was this page helpful?