The Giveaway and the Lockout: The Week the AI Frontier Split in Two

On Monday, a company most people cannot name gave away a frontier model. That same day, the Pentagon stopped using Anthropic's. Two moves in opposite directions. Together, they redraw where the power in AI sits.

Reflection published Beam: a 501-billion-parameter model, pretrained on 23.8 trillion tokens and trained with more than 100 million reinforcement-learning rollouts across 10,500 of Nvidia's newest GB300 chips. The weights are due for free release under Apache 2.0 this month.

That same day, the Pentagon confirmed the opposite. It has stopped using Anthropic's AI. Claude, built by the lab that made safety its brand, is out of the US military's toolchain, months after the department labelled the company a national-security risk.

One frontier was handed out. Another supplier was shut out. Neither makes sense without the other.

The giveaway

Beam is a sparse mixture-of-experts model: 501 billion parameters, 23 billion active per token. Reflection trained it over four weeks. It reports 80.9 on SWE-Bench Verified and 80.1 on Terminal-Bench v2.1, matching larger open models such as GLM-5.2 while using three to four times less compute at inference.

Those numbers are self-reported. The licence is the point: the weights and the full stack for running and fine-tuning Beam will land under Apache 2.0, which permits commercial use and redistribution. That is the most permissive common licence, and the opposite of a moat.

Beam did not arrive alone. DeepSeek was reported to be raising about $12 billion from Tencent and CATL, ahead of a Hong Kong listing targeted for early 2027. Etched, an inference-chip startup, was fielding offers valuing it at $40 billion to $50 billion, double its August mark. The money is not betting on scarcity. It is betting on volume.

The lockout

In February, according to the BBC, the Pentagon asked Anthropic to strip the guardrails from Claude and give the military unfettered access. Anthropic refused, citing mass surveillance and autonomous weapons. Defence Secretary Pete Hegseth then designated the company a national-security supply-chain risk, a label usually reserved for hostile states. Anthropic called it unlawful and sued.

By late August the Pentagon was supposed to have stopped using Claude. It had not. Sources told the BBC that as recently as last week, Claude was still doing intelligence work for the department, including in operations against Iran, much of it through Palantir's Maven system. On Monday, an official said the phase-out was complete.

The reasons have not been published. The sequence has: the lab that argued for caution lost its place after refusing to drop it. And heise reports that Meta and Microsoft are also restricting internal use of Anthropic's models.

What the two stories share

Capability is becoming a public good: a frontier model can be trained, then given away, and the price of that intelligence falls toward zero. Buyers, meanwhile, are sorting suppliers by trust, and trust is being withdrawn from the lab that made caution its selling point.

For three years the assumption has been that the best weights are the moat. Beam punctures that. If a 501-billion-parameter model can be released for anyone to rebuild, the model is not the advantage. And responsibility, the second supposed advantage, takes its own hit: the Pentagon did not reward Anthropic's caution, it removed the company for it.

Beam proved a brain can be copied. The Pentagon proved trust cannot. Only one of those is for sale.

The honest counterpoint

Both moves have a serious defence. Open weights are a strategy, not charity: China proved they win users, and a Western lab matching them is a fair answer. The Pentagon's caution is not invented either: an unfiltered model aimed at targeting raises exactly the questions Anthropic refused to wave away.

But the defences do not cancel the pattern. They sharpen it. Give your frontier away, and remove the cautious vendor, in the same week, and you have told every builder the same thing: the layer you do not hold can be given to everyone else, or taken from you, and neither decision is yours.

What it means if you build on AI

I build local and on-prem systems for a living, so I will not pretend to be neutral. The lesson is not to pick a side between labs and governments. It is the one this journal keeps arriving at. Own the layers you depend on. Keep a second provider, so one enforcement decision is an inconvenience, not an outage. Keep your work where you can export it. And run what you can run yourself for the work that cannot afford a gap.

Beam is the hopeful half of the week. The Pentagon story is the caution: holding the weights is not enough if the relationship around them can be cut by a decision you never see coming. The safest AI is still the one on your own machine, from weights you actually have.

One caveat I cannot settle. If Reflection's benchmarks do not survive independent testing, or if the Pentagon and Anthropic reconcile, this reading weakens. I will update the piece if either happens.

Readout: what is confirmed, and what is not

• Confirmed (Reflection, 5 Oct 2026): Beam is a 501B-parameter sparse MoE with 23B active, pretrained on 23.8T tokens and trained with 100M+ RL rollouts on 10,500 Nvidia GB300 GPUs; weights and stack are slated for Apache 2.0 release this month.
• Self-reported (Reflection): 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench v2.1, and parity with GLM-5.2 at 3 to 4 times lower inference compute. Awaiting independent verification.
• Confirmed (BBC, 5 Oct 2026): a DoD official said the Pentagon "has ceased the use of Anthropic products".
• Reported (BBC, 5 Oct 2026): multiple people said Claude was still in use last week for research, analysis, intelligence and operations against Iran, including via Palantir's Maven. Anthropic has not confirmed this.
• Reported (BBC, Guardian, Feb 2026): the Pentagon asked Anthropic to remove guardrails; Anthropic refused; Hegseth designated it a supply-chain risk; Anthropic sued.
• Reported (heise, 6 Oct 2026): Meta and Microsoft are restricting internal use of Anthropic's models.
• Reported (Bloomberg, CNBC, 6 Oct 2026): DeepSeek is raising at least $12B from Tencent and CATL, eyeing an early-2027 IPO. Reported (TechCrunch, 5 Oct 2026): Etched is fielding $40B to $50B offers.

If you are working out which layers of your own stack you actually hold, and what it would take to keep them, email brandon@kreostudio.co.uk.

Sources & references

1. Introducing Beam: Reflection's 501B open-weight model, Reflection AI, 5 October 2026.
2. Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost, TechCrunch, 5 October 2026.
3. Pentagon stops using Anthropic AI tools after blacklisting company, BBC News, 5 October 2026.
4. Anthropic says it cannot in good conscience allow Pentagon to remove AI checks, The Guardian, 26 February 2026.
5. Anthropic: Pentagon completes phase-out, Meta and Microsoft restrict usage, heise online, 6 October 2026.
6. DeepSeek to raise at least $12 billion in Tencent-backed funding, Bloomberg, 6 October 2026.
7. DeepSeek latest funding round could reach $15 billion, CNBC, 6 October 2026.
8. Etched fields funding offers at $40B+ valuation, TechCrunch, 5 October 2026.