Anthropic researchers had GLM-5.2 complete a full replication chain on all eight $55 TV-box test nodes in four of five runs. They say no current model is both small and capable enough to spread in the wild, and that gap could close by the end of 2026 at the current pace.
Perspective Coverage
11 publishers
- Builder
- Builder 34%
- Operator
- Operator 45%
- Investor
- Investor 21%
Reality
- Evidence66
- Adoption
- Not scored
- Hype gap+32
- Incentives
- Not scored
- Confidence72
Anthropic ran Gemma4-26B-A4B across 15 off-brand Android TV boxes at about $55 each to test whether an AI worm could host itself on IoT gear. Its researchers theorize a working worm could follow by year end. Unmanaged devices need an owner now.
Perspective Coverage
5 publishers
- Builder
- Builder 34%
- Operator
- Operator 45%
- Investor
- Investor 21%
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+35
- Incentives50
- Confidence60
buildOne report1 publisher Coinbase trained its 9B fraud reviewer on a reward computed from known labels, with no judge model
Coinbase says a Qwen3.5-9B model post-trained for under $100 beat Claude Opus 4.5 on its Onramp fraud benchmark. The figures are Coinbase's on private data; the reusable parts are the label-based reward and the tier-only, block-only harness.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+25
- Incentives55
- Confidence55
buildConfirmed7 publishers Microsoft has put Decision-1, a model that scores fixed choices, in Foundry at $0.042 per million input tokens. Agent teams now have to test whether its scores, so far benchmarked only by Microsoft, are calibrated well enough to decide when a case goes to a human.
Perspective Coverage
7 publishers
- Builder
- Builder 42%
- Operator
- Operator 28%
- Investor
- Investor 30%
Reality
- Evidence38
- Adoption
- Not scored
- Hype gap+30
- Incentives65
- Confidence62
buildOne report1 publisher Self-edited context lifts Qwen3.5-9B from 28.8% to 42.5% on BrowseComp-Plus
Meta, MIT and University of Washington researchers let models rewrite their own context, lifting a trained 9B model from 28.8% to 42.5% on BrowseComp-Plus. The gains come from specified benchmark tasks, and rewriting earlier context invalidates the model cache, so we would keep hand-built compaction pipelines as the production default.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+20
- Incentives
- Not scored
- Confidence40
Newer AI models caught less fraud in Coinbase's replay of 16,140 real transactions
Coinbase replayed 16,140 real transactions and found each newer version of three AI model families caught less fraud than the one it would replace. The result undercuts the idea that a model upgrade is a free improvement for a working fraud screener.
Reality
- Evidence50
- Adoption40
- Hype gap+15
- Incentives50
- Confidence50
Microsoft claims its Decision-1 scoring model answers 35 times faster than GPT-6 Sol
Microsoft released Decision-1, a model that scores fixed answer options and, by its own benchmarks, runs 35 times faster than GPT-6 Sol. Whether routing work moves off general-purpose models now depends on independent tests and on pricing that confirm those self-reported figures.
Reality
- Evidence30
- Adoption15
- Hype gap+35
- Incentives65
- Confidence35
Microsoft's Agent Lightning trains AI agents inside the harness they ship with
Microsoft Research Asia open-sourced Agent Lightning v1.0, a 3,500-line framework whose recipe lifted a 9B model's SWE-bench Verified score by 14.6 points. It runs reinforcement learning on the agent a team already deploys, sitting as a proxy between that agent and its model.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+15
- Incentives60
- Confidence40
buildOne report1 publisher UW and Meta's Context Language Models beat Codex-style summarisation by six points on BrowseComp-Plus
UW and Meta researchers report 59.4% on BrowseComp-Plus for a model that edits its own context, against 53.4% for Codex-style summarisation. The edited file is thrown away when the task ends, so memory that lasts across tasks is still the builder's job.
Reality
- Evidence42
- Adoption8
- Hype gap+12
- Incentives
- Not scored
- Confidence45
buildOne report1 publisher Ollama's new endpoint puts typed routing and moderation decisions on the local machine
Ollama 0.35 adds a /v1/systemone endpoint returning a choice, yes/no or score from models run on the device. Its 9B Nimble model matched human moderation labels 70.3% of the time in its maker's test, so each team still sets its own review threshold.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+20
- Incentives60
- Confidence45
buildOne report1 publisher Probes on a 27B open model match direct probes of a 397B model on deception
Probes on Qwen3.5-27B reading other models' text came within 0.004 AUROC, on average, of probing authors up to 397B directly, a LessWrong post reports. Every tested pair was open-weight, so auditors who apply the method to closed models get the reader's view of the text and cannot measure that gap.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+20
- Incentives
- Not scored
- Confidence30
buildOne report1 publisher Qwen3.5-9B's 262K context window would consume the whole 8GB budget in KV cache
A dev.to guide to running local models on 8GB prices the KV cache between 15KB and 160KB per token depending on architecture. At 32K tokens held, that spread is the difference between 0.5GB and 5GB of a fixed budget.
Reality
- Evidence22
- Adoption
- Not scored
- Hype gap+45
- Incentives62
- Confidence58
Harvey post-trained a 27B open-weight model into the frontier band on its own legal benchmark
The band it joined is one where the best closed models still finish under 10% of tasks end to end at roughly $50 and 20 minutes each, so what post-training buys a law firm is price and hosting rather than capability.
Publishers:harvey.ai
Reality
- Evidence41
- Adoption
- Not scored
- Hype gap+32
- Incentives76
- Confidence46
buildOne report1 publisher A small harness carried a local model through the regression check in 2.1 minutes
A June finding said local models cannot iterate on code. It said so about chat UIs, and it named the fix. July supplied that fix and measured it. The follow-up test result is the number worth reading.
Reality
- Evidence46
- Adoption12
- Hype gap+12
- Incentives32
- Confidence44
buildOne report1 publisher Agent Lightning v1.0 hands the RL loop to the harness, and the trainer gets text, not tokens
Microsoft reports Qwen3.5-9B going from 41.8% to 56.4% on SWE-bench Verified using 6K examples. The load-bearing change is who owns the agent loop, and what the service boundary costs.
Reality
- Evidence32
- Adoption18
- Hype gap+28
- Incentives66
- Confidence44
buildOne report1 publisher Same weights, 70 points apart: the ARC-AGI-3 table has stopped being procurement evidence
One model was verified at 30.16 in the official harness and reported at 100.00 in NVIDIA's. Microsoft's Agent Lightning now trains the harness into the weights.
Reality
- Evidence54
- Adoption58
- Hype gap+28
- Incentives66
- Confidence48
buildOne report1 publisher A 170-goal agent field test costs $0.49. Proving it actually passed costs more.
PlannerCritic's sweeps are cheap enough to gate every release. The harder problem is telling a clean run from a rig that quietly counted nothing.
Reality
- Evidence34
- Adoption11
- Hype gap+24
- Incentives68
- Confidence33