Issue 17 · Week 33 of 2026.

5 model rows added and 1 updated — the heaviest single week the tree has recorded in 2026. Two of the five additions are classified gated rather than open_weights despite open-source framing in their own announcements, and the updated row publishes weights for a build its API no longer serves.

The openness classification did more work this week than the branch classification. GLM-5.3 and GPT-5.6-Cyber are both recorded as gated: one because its weights are promised but unshipped, the other because access runs through a vetting tier with hardware security keys. Reclassify GLM-5.3 to open_weights when the Hugging Face release is confirmed — the tree records what shipped, not what was announced.

Vendor-stated frontier capability. The releases that reset the closed-source ceiling.

• 2026-08-12/SpaceXAI/Frontier/Reasoning

  Grok 4.6

  Ties GPT-5.6 Sol Max at 61 on the Intelligence Index at $2 / $6, and resolves long-horizon agentic tasks in roughly half the turns of Claude Opus 5

  Explicitly a post-training release on the Grok 4.5 foundation rather than a bigger base model, five weeks after its predecessor. The tie on the Index is the least useful number in the release: architects should look at the measured ~53 turns and ~0.5B input tokens against Claude Opus 5's ~103 turns and ~2.0B at max settings, because turn count drives the bill on any long-running agent and does not move when a rate card changes. The gap it did not close is Terminal-Bench v3.0, where 26% against GPT-5.6 Sol Max's 34.6% still argues against putting it on unattended terminal work.

  VentureBeat, Artificial Analysis, The Agent Report

• 2026-08-13/Google/Frontier/Multimodal

  Gemini 3.7 Flash

  1M context workhorse at $0.75 / $3.75, half of 3.6 Flash — with a doubling to $1.50 / $7.50 published at launch for January 1

  Three weeks after 3.6 Flash, which is a cadence that tells you where Google is putting its attention: the Pro line has no announced timeline and Pichai declined Pro-cadence questions on the last earnings call. Buyers should treat the price as the story and read the footnote — Google's own pricing page states the rate is introductory through December 31. This is the clearest public admission by any lab that current agent unit economics are promotional, and it is disclosed rather than discovered, which is the honest way to do it.

  Google blog, Google Cloud pricing page, InfoWorld

• 2026-08-10/OpenAI/Specialist/Reasoning

  GPT-5.6-Cyber

  Purpose-trained security model gated behind a new Daybreak Red tier at $12.50 / $75, with hardware security keys required from September 1

  Reported to answer 95.0% of advanced cyber requests against 1.5% for GPT-5.6 Sol under normal safeguards, and rated High rather than Critical under OpenAI's Preparedness Framework. The release is a distribution decision more than a capability one: Daybreak Blue gives approved defenders general-purpose models with cyber guardrails removed, Red gates the purpose-trained models behind vetting. Security leaders should read this as the first frontier model where the access-control design is the product, and plan for the vetting timeline rather than the API key. Recorded at medium placement confidence: no first-party OpenAI source was located during research.

  AI/TLDR, DEV Community, The Neuron

Open-weights releases that change procurement options. Pull these into pilot when score parity meets license parity.

• 2026-08-10/Meta/Edge / small/Dense

  Muse Glimmer

  30B dense multimodal under Apache 2.0 — Meta's first open release with no user gate, no naming rule and no acceptable use policy

  Artificial Analysis scores it 35 on the Intelligence Index and 44 on the Openness Index, level with DeepSeek V4 Flash and GLM-5.2. The capability is mid-tier for its class; the license is the release. Every prior Meta open model carried a 700-million-user gate and a 'Built with Llama' naming requirement, and Apache 2.0 leaves Meta no lever to withdraw the grant — so teams that rejected Llama on legal review should re-run that review. Two caveats worth carrying into the decision: Meta withheld training data and code, and it serves no API for the model, so every price and latency figure comes from a third party.

  Meta AI Research, Artificial Analysis, VentureBeat

• 2026-08-12/DeepSeek AI/Open frontier/MoE

  DeepSeek-V4 Pro (0813)

  Flagship leaves preview with large vendor-reported gains — announced by nothing but a version string, and with the April weights still the only ones downloadable

  Terminal Bench 2.1 from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, CyberGym from 52.7 to 83.3, Intelligence Index from 45 to 53 — all vendor-reported, all on a build that was not on Hugging Face when the endpoint swapped. Buyers running V4 Pro through the API got a materially different model this week without a release note; buyers self-hosting got nothing. Anyone with a benchmark result for 'DeepSeek V4 Pro' in a procurement document should date it and note which build it refers to, because the name now covers two materially different artifacts.

  Unite.AI, Decrypt, OpenLLMStack, Oflight

• 2026-08-14/Z.ai/Open frontier/MoE

  GLM-5.3

  Same 753B MoE as GLM-5.2 with a long-horizon post-training run — best open-source Terminal Bench 3.0 claimed, weights promised within two weeks

  The architecture did not change at all; the delta is post-training inside sandboxes built to mimic developer workstations, with some tasks running for days. That is a useful data point on how much long-horizon capability is now recoverable from post-training alone on a fixed base. Z.ai reports a CyberGym score ahead of Claude Mythos 5 while trailing it on two other cybersecurity benchmarks, which is a narrower claim than the headline suggests. Recorded as gated: at announcement the only access was a paid coding subscription, so treat the open-source framing as an intention until the Hugging Face release lands.

  SiliconANGLE, Z.ai

Architectural patterns that crossed multiple vendors this period. Each pattern lists exemplar releases and what it changes for deployment, cost, or capability.

• Post-training is the release; the base model stopped moving

  Grok 4.6GLM-5.3DeepSeek-V4 Pro 0813

  Three of this week's six releases explicitly ship the same foundation as their predecessor with a longer or differently-shaped post-training run. SpaceXAI said so directly for Grok 4.6 — regenerated supervised fine-tuning trajectories plus reinforcement learning in agentic environments — and Z.ai's GLM-5.3 is architecturally identical to GLM-5.2 at 753B MoE and 1M context. The procurement consequence is that version numbers no longer imply retraining cost or a capability step change, and the release cadence a vendor can sustain is now decoupled from its pretraining compute. Architects sizing a model refresh cycle should stop assuming a new point release means a new base.

  The Agent Report, SiliconANGLE, Oflight

• The served build and the published weights have separated

  DeepSeek-V4 Pro 0813GLM-5.3Muse Glimmer

  DeepSeek swapped its production endpoint to a new build while Hugging Face continued to host the prior one. Z.ai benchmarked a model available only by subscription and promised weights later. Meta published weights for a model it does not serve at all, so the artifact is downloadable but has no reference endpoint. In each case the name on the benchmark and the name on the download resolve to different things, which breaks the assumption underneath every open-weight evaluation: that a published score describes an artifact you can obtain. Combined with last week's Endpoint Accuracy Index finding of up to 40% accuracy loss by serving configuration, a model score is now a claim about a build-and-endpoint pair, not about a model.

  Unite.AI, SiliconANGLE, Artificial Analysis

• Published price expiries replace silent repricing

  Gemini 3.7 FlashDeepSeek-V4 ProGPT-5.6-Cyber

  Google states on its pricing page that Gemini 3.7 Flash's $0.75 / $3.75 runs through December 31 and becomes $1.50 / $7.50 on January 1. DeepSeek warned of a significant increase and then landed peak and off-peak tiers effective August 16, reported at $1.32 / $3.96 peak against a flat $0.435 / $0.87 before. Both are disclosed in advance rather than discovered in an invoice, which is a genuine improvement in vendor behavior — and also the clearest signal yet that the current cost of inference is subsidized. Finance teams should build the published step-up into any agent business case whose payback period crosses January.

  Google Cloud pricing, OpenLLMStack, InfoWorld

• Dense at 30B as the local-agent target, against the MoE consensus

  Muse GlimmerLFM2.5-2.6BQwen3.6-27B

  Nearly every frontier release this year has been a mixture-of-experts design, and Muse Glimmer is deliberately not: 30B parameters that all activate on every forward pass, shipped with two 4-bit quantizations and a speculative-decoding drafter at a roughly 20 GB deployment envelope. Dense is the easier target for consumer runtimes and quantization toolchains, which is the point when the deployment surface is a single 24-32 GB consumer GPU rather than a datacenter. The pattern to watch is a bifurcation: MoE where memory bandwidth is cheap and dense where the constraint is a laptop.

  Meta AI Research, MindStudio, Artificial Analysis

Benchmark deltas that change a procurement read. Scores reflect public leaderboards or vendor model cards as of publication.

• Artificial Analysis Intelligence Index

  Grok 4.6 enters at 61 on a post-training release alone, tying GPT-5.6 Sol Max and putting four models within two points of the lead

  • Claude Opus 563
  • Claude Fable 562
  • Grok 4.661
  • GPT-5.6 Sol Max61
  • DeepSeek-V4 Pro 081353 (vendor-reported, up from 45)

  Artificial Analysis via VentureBeat and The Agent Report

• GDPval-AA v2 (real-world knowledge work, Elo)

  Grok 4.6 takes the top position at 1,753, a 227-point jump over Grok 4.5 and the largest single-release move on this board this year

  • Grok 4.61,753
  • Claude Fable 5 Max1,741
  • GPT-5.6 Sol Max1,728
  • Grok 4.51,526

  Artificial Analysis via VentureBeat

• CyberGym (vulnerability discovery)

  Two separate vendors claimed the top spot on the same board in the same week, one open-weight and one Chinese subscription-gated — and neither claim has been independently reproduced

  • DeepSeek-V4 Pro 081383.3 (vendor-reported)
  • Claude Fable 583.1
  • Claude Opus 4.878.3
  • GLM-5.3Ahead of Claude Mythos 5 (vendor-reported, figure not disclosed)

  DeepSeek via Oflight and Decrypt; Z.ai via SiliconANGLE

• Terminal-Bench (v2.1 and v3.0 — not comparable to each other)

  The board split into two versions being quoted interchangeably, which is where most of this week's coverage went wrong

  • DeepSeek-V4 Pro 0813 (v2.1)87.9 (vendor-reported, up from 72.1)
  • Claude Fable 5 (v2.1)88.0
  • GPT-5.6 Sol Max (v3.0)34.6%
  • Grok 4.6 (v3.0)26%
  • GLM-5.3 (v3.0)Highest open-source score claimed (vendor-reported)

  Oflight, VentureBeat, SiliconANGLE

• Vals Index, open-weights view

  DeepSeek-V4 Pro 0813 moves to second at 66.25% from the April build's 55.62%, and the cost column separates it further than the score does

  • DeepSeek-V4 Pro 081366.25% at $0.14 per test
  • Kimi K3Behind on score, $2.34 per test
  • Qwen3.8 MaxBehind on score, $2.68 per test
  • DeepSeek-V4 Pro (April)55.62%

  Vals AI via OpenLLMStack

A snapshot of leader-vs-challenger by tier. Useful for procurement shortlists when matching workload to model class. Pair with the benchmark moves above for the underlying scores.

• Closed frontier

  Leader: Claude Opus 5

  Challenger: Grok 4.6

  Opus 5 holds the Intelligence Index lead at 63, but Grok 4.6 tied GPT-5.6 Sol Max at 61 this week on a post-training run alone while charging $2 / $6 against Sol's $5 / $30. Four models now sit within two points, so the tier is decided on cost per completed task rather than on the Index.

• Open frontier

  Leader: DeepSeek-V4 Pro 0813

  Challenger: Kimi K3

  Leader on served scores and unbeatable on price at $0.14 per Vals test, with an asterisk this publication is applying for the first time: the 0813 weights were not downloadable when the endpoint swapped, so the leading open model was briefly not obtainable. GLM-5.3 may take this row within two weeks if its promised weights land.

• Reasoning

  Leader: Claude Opus 5

  Challenger: GPT-5.6 Sol Max

  Unchanged on capability, but the efficiency read moved against Opus 5 this week: Artificial Analysis measured it taking roughly 103 turns and 2.0B input tokens on long-horizon agentic tasks against Grok 4.6's ~53 turns and ~0.5B for a comparable answer quality.

• Coding

  Leader: Claude Fable 5 Max

  Challenger: GPT-5.6 Sol Max

  Fable 5 Max leads CursorBench v3.2 at 70.5% and FrontierCode v1.1 Extended at 63.6%, while GPT-5.6 Sol Max takes DeepSWE v1.1 at 73%. Grok 4.6 closed most of the gap on DeepSWE this week, rising from 54% to 65.9%, without taking either row.

• Multimodal

  Leader: Gemini 3.7 Flash

  Challenger: Muse Spark 1.2

  Google took this row on economics rather than capability: 1M context with text, image, audio and video input at $0.75 / $3.75 is half the prior Flash rate, and the model went straight into Gemini Spark. Read the January 1 doubling to $1.50 / $7.50 before treating it as the durable price.

• Edge / small

  Leader: Muse Glimmer

  Challenger: Qwen3.6-27B

  Muse Glimmer takes the row on licensing, not on score — 35 on the Intelligence Index is mid-pack for its class, but Apache 2.0 with no user gate, naming rule or acceptable use policy removes the legal friction that kept Llama out of many production pipelines. Liquid's LFM2.5-2.6B remains the pick well below 20 GB.

The non-release moves that shift vendor risk — pricing, deprecations, gating decisions, license changes — with a one-line procurement read.

• 2026-08-12/DeepSeek AI

  Flagship endpoint swapped to the 0813 build with no announcement, and a price increase scheduled for August 16 replacing the flat rate with peak and off-peak tiers

  Two procurement problems in one move. The silent build swap means any team with a pinned expectation of model behavior got a different model without notice, so anyone running DeepSeek in production should re-run their own evals this week. The pricing change, reported at $1.32 / $3.96 peak against $0.435 / $0.87 before, lands the day after this issue publishes and materially changes the cost case that made V4 Pro the default cheap frontier option.

  Unite.AI, Decrypt, OpenLLMStack, Oflight

• 2026-08-10/Meta

  First open release under Apache 2.0, abandoning the Llama License's user gate, naming rule and acceptable use policy, with Muse Spark 1.2 weights promised to follow

  Legal and procurement teams that blocked Llama-family models on license terms have lost their objection for this model, and should re-run the review rather than carry forward a stale decision. The strategic read is that Meta stopped charging in terms — a price it could only charge while buyers had no substitute — and the promised Muse Spark 1.2 weights would extend that to a frontier-class model.

  Meta AI Research, VentureBeat, Business Model Analyst

• 2026-08-10/OpenAI

  Daybreak program split into Blue and Red tiers, with hardware security keys required for login from September 1

  The first frontier lab to make physical authentication a condition of model access. Security teams planning to use Daybreak-tier models for vulnerability research should start the vetting and key-provisioning process now rather than at the point of need, because the gate is organizational rather than technical and the September 1 date is firm.

  AI/TLDR, DEV Community

• 2026-08-13/Google

  Third Flash release in roughly six weeks while the Pro line has no announced timeline and Pichai declined Pro-cadence questions on the last earnings call

  Google is iterating fast where the volume is and slowly where the headlines are. Architects should not wait on a Pro refresh to make agent platform decisions this quarter, and should note that the Flash tier now carries Pro-level agentic positioning in Google's own documentation.

  Google blog, InfoWorld

• 2026-08-14/Z.ai

  GLM-5.3 benchmarked and announced as open-source with weights promised on Hugging Face within two weeks, available at launch only through a paid coding subscription

  A benchmark-first, weights-later release pattern that is becoming common enough to plan around. Teams should not schedule a self-hosted deployment against an announcement date, and should treat the two-week window as a forecast rather than a commitment — the tree will reclassify the model only when the weights are confirmed.

  SiliconANGLE

Specific model-side catalysts in the next 7–30 days that would change the read materially. Watching these tells us whether the canopy is widening or thinning.

• Aug 16

  DeepSeek peak / off-peak pricing takes effect

  The flat $0.435 / $0.87 rate ends and reported peak tiers of $1.32 / $3.96 begin, with peak hours at 01:00-04:00 and 06:00-10:00 UTC. Watch whether workloads visibly shift into off-peak windows, which would be the first evidence of time-of-day arbitrage becoming a real inference architecture pattern.

• Aug 16-28

  GLM-5.3 weights on Hugging Face

  Z.ai promised weights within two weeks. If they land, the model likely takes the open frontier row and this publication reclassifies it from gated. If they slip, the benchmark-first release pattern is worth treating as a durable vendor behavior rather than a one-off.

• Aug 16-31

  DeepSeek-V4-Pro-0813 weights, and independent reproduction of its gains

  Every headline number for the 0813 build is vendor-reported. The Terminal Bench 2.1 jump from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7 are large enough that independent replication is the only thing that should move them into a procurement document.

• Sep 1

  OpenAI hardware security key requirement for Daybreak login

  The first hard physical-authentication gate on frontier model access. Whether other labs follow within the quarter determines if this becomes a category norm for dual-use capability or stays an OpenAI-specific control.

• Aug 16-Sep 15

  Muse Spark 1.2 open weights

  Zuckerberg said the weights are coming. A frontier-class Meta model under a permissive license would be the largest single addition to the open tier this year and would force a re-read of the open-versus-closed capability gap lever.

• Sep-Oct

  Independent turn-efficiency measurement across the frontier

  Grok 4.6's ~53-turn result against Claude Opus 5's ~103 is currently a single-source measurement on a small set of tasks. If Artificial Analysis or another independent lab extends turn counting across the full frontier, it becomes the most decision-relevant number in agent procurement.