AI Signal — 2026-10-28

AI Field Status

The industry's center of gravity has moved from frontier capability to cost per completed task. Frontier closed models (Opus 5.5, Fable 5.1) still lead on hard long-horizon work, but open-weight Chinese models (DeepSeek, Kimi, GLM) now sit within a small quality gap, at prices two to three orders of magnitude lower. Advantage now comes from routing, inference hosting and evaluation discipline, not from which single model a team picks.

Today's Thesis

Near-frontier agentic capability is now open-weight and almost free per task, so enterprise AI strategy shifts from choosing a model to routing workloads across closed, hosted-open and self-hosted tiers based on your own evals.

Key Takeaways

Default model selection is no longer defensible. A quoted cost per task of $0.26 versus $211–$763 means any agentic workload without a routing layer is overpaying by orders of magnitude.
Cache-hit pricing (free on Lithos, $0.01/M on Base10) makes loops with long repeated context nearly free, so prompt and context architecture is now a direct cost lever.
Sovereign and on-prem agentic coding is now credible. A 50-minute autonomous build ran fully on one DGX Station, which strengthens the Private AI case for regulated clients.
Open weights expose full reasoning and tool-call traces. That gives an auditability edge closed APIs can't match, which matters for compliance-heavy deployments.
The evidence is one sponsored, single-run video with secondhand benchmark figures. Treat it as a hypothesis to test on internal evals, not a migration trigger.

Executive Signal Scoring

Most Important

Cost per task has collapsed for near-frontier work: open-weight models now sit on the quality-per-cost Pareto frontier, priced several orders of magnitude below closed flagships.

Most Actionable

This week, run DeepSeek V4.1 Flash through a hosted endpoint against your top three agentic workloads on internal evals, and measure task success and cost per completed task against your current default.

Most Overhyped

The claim that 'DeepSeek crushed everyone.' By the video's own numbers it trails Opus 5.5 and only ties Kimi K3 and GLM 5.3 at about 42 on Humanity's Last Exam and Terminal-Bench, and the demos are cherry-picked single runs.

Biggest Blind Spot

Most enterprises lack an internal eval harness and routing layer. Without them they can't capture the cost collapse safely, so they either overpay for closed models or adopt open models without evidence of quality and supply-chain risk.

Most Likely Next Shift

Inference hosts and routing or gateway layers will capture margin from model vendors. Closed frontier labs will respond with aggressive price cuts or caching discounts on their mid-tier models.

Signal Note

What Landed

Matthew Berman argues that DeepSeek V4.1 Flash, an open-weights multimodal model, offers the best mix of quality, cost and speed available. He concedes that Opus 5.5 still beats it on harder builds such as the Fall Guys clone. His main number comes from Artificial Analysis: about $0.26 per completed task, against $211 for Kimi K3.1 Pro and $763 for Claude Fable 5.1 Max. He also ran a 50-minute autonomous coding job entirely on a local DGX Station, and he quotes hosted inference at 1,286 tokens/s on Lithos ($0.15 per million input tokens, $0.60 per million output, cache hits free).

Why It Matters

If the per-task figures hold up on a buyer's own tests, sending every agentic workload to a premium closed model by default becomes hard to justify. The cost levers become routing, caching and the choice of host. Free cache hits help agent loops that resend the same long context most. A local run with full reasoning and tool-call traces strengthens the case for private or sovereign deployment. However, benchmark parity is only with Kimi K3 and GLM 5.3, at about 42 on Humanity's Last Exam and Terminal-Bench, not with the frontier. The evidence is single-run, cherry-picked demos in a CoreWeave-sponsored video, and the pricing is secondhand.

Worth Raising With Customers

Tiered routing: send high-volume, well-specified agent tasks to an open-weights tier and keep premium closed models for the hardest reasoning. Berman's own PDF-reader build worked this way: Opus 5.5 wrote the spec and the cheap model built it.
Cost per completed task: before any migration, run the model against the customer's own eval set and compare cost per completed task, not price per token.
Regulated or sovereign buyers: running open weights on owned hardware is now a credible option for coding agents. The tested setup was a DGX Station, though. Quantized versions on consumer GPUs are untested.

Sources

Expert	Source	Published	Source text	Summary
Matthew Berman	Deepseek crushed everyone	2026-10-28	ok	ok