--- title: >- Model governance now starts at the deployable package: entitlement, retention, runtime version, and price window publication: The Model Pulse slug: 2026-W36 issueNumber: 20 isoYear: 2026 isoWeek: 36 cadence: weekly publishedAt: '2026-09-05' periodLabel: Week 36 of 2026 canonicalUrl: https://brianletort.ai/industry/models/2026-W36 pdfUrl: https://brianletort.ai/downloads/model-pulse-2026-W36.pdf schemaVersion: 2026.05.02 treeDelta: added: - gpt-6-astra - claude-fable-5-1 - gemini-3-8-flash updated: [] frontierMovements: - modelId: gpt-6-astra name: GPT-6 Astra vendor: OpenAI releaseDate: '2026-09-03' tier: frontier architecture: agentic - modelId: claude-fable-5-1 name: Claude Fable 5.1 in GitHub Copilot vendor: Anthropic releaseDate: '2026-09-01' tier: frontier architecture: reasoning - modelId: gemini-3-8-flash name: Gemini 3.8 Flash in GitHub Copilot vendor: Google releaseDate: '2026-09-03' tier: frontier architecture: reasoning openWeights: - modelId: muse-glimmer-30b name: Muse-Glimmer-30B NVFP4 vendor: NVIDIA releaseDate: '2026-09-01' tier: open_frontier architecture: multimodal - modelId: null name: NeoMME-800M vendor: H Company releaseDate: '2026-09-03' tier: specialist architecture: multimodal architecturePatterns: - Runtime version becomes a governed SKU property - Distribution policy is part of the model SKU - Open-weight progress shifted from new lineage to deployment compression benchmarks: - ARC-AGI-3 Semi-Private - Artificial Analysis Intelligence Index v4.1.1 - Terminal-Bench 2.1, vendor-reported checkpoint comparison scorecardAsOf: '2026-09-05' vendorSignals: - vendor: OpenAI date: '2026-09-03' signal: >- Astra enterprise access off by default; Standard API priced at $10/$50 per million input/output tokens - vendor: Anthropic / GitHub date: '2026-09-01' signal: >- Fable 5.1 reaches general availability but Business and Enterprise administrators must enable it - vendor: Google / GitHub date: '2026-09-03' signal: Gemini 3.8 Flash enters Copilot under introductory provider pricing through December 31 - vendor: NVIDIA date: '2026-09-01' signal: Muse-Glimmer-30B NVFP4 ships Apache 2.0 for Blackwell with a 24.7 GB checkpoint --- # Model governance now starts at the deployable package: entitlement, retention, runtime version, and price window *The Model Pulse · Issue 20 · Week 36 of 2026 · Published 2026-09-05* ## Big Read The differentiated procurement move this week is not the already-circulating claim that the harness matters. It is that a model stock-keeping unit (SKU) now includes administrator enablement, retention route, provider-adapter version, reasoning effort, and promotion expiry. Claude Fable 5.1 is broadly available in GitHub Copilot but administrator-enabled for enterprises and follows a documented default-retention route outside approved zero-data-retention arrangements. Gemini 3.8 Flash enters the same surface under introductory provider pricing, while Astra enterprise access is off by default. ARC Prize supplies the evidence for why runtime version belongs in that record: its controlled maximum-reasoning comparison produced a 35.9-point harness-associated spread. The best-observed adapter spread is larger but changes reasoning effort, so it is descriptive rather than causal. This launch-week finding belongs to ARC Prize and prior coverage; the operational extension here is to version the whole deployable SKU and rerun it without provider-private state before claiming portability. Open weights were quiet after W35's five-model wave. NVIDIA's Apache-2.0 Muse-Glimmer-30B low-precision checkpoint reduced storage substantially, but it is a quantization update to an existing model, not evidence that the open frontier closed the gap this week. The model tree adds Astra, Fable 5.1, and Gemini 3.8 Flash without manufacturing a new open-frontier leader from a footprint optimization. ## Tree delta 3 model rows added: GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash. No existing model row changed; Muse-Glimmer's low-precision checkpoint is an update, not a new lineage node. **Added (3):** `gpt-6-astra`, `claude-fable-5-1`, `gemini-3-8-flash`. **Updated:** none. *W36's delta is closed and distribution-heavy. The durable change is that runtime state, administrator entitlement, retention, and promotional pricing now sit beside architecture in model diligence.* ## Frontier movements ### GPT-6 Astra `gpt-6-astra` **OpenAI** · 2026-09-03 · frontier · agentic _1M-context closed agent model with a 35.9-point matched-effort harness-associated ARC-AGI-3 spread_ Evaluate Astra as a model-plus-runtime release. ARC Prize's matched-effort comparison associates a 35.9-point spread with the harness; portability reviews should reproduce it without provider-private reasoning state. Source: [OpenAI, ARC Prize](https://arcprize.org/blog/astra). ### Claude Fable 5.1 in GitHub Copilot `claude-fable-5-1` **Anthropic** · 2026-09-01 · frontier · reasoning _General availability across Copilot surfaces with administrator enablement and retention conditions_ Engineering leaders get a broadly distributed coding model, not an automatically enabled enterprise SKU. Put administrator entitlement and the documented 30-day default retention route into the architecture record alongside model version and benchmark results. Source: [GitHub](https://github.blog/changelog/2026-09-01-claude-fable-5-1-generally-available-in-github-copilot/). ### Gemini 3.8 Flash in GitHub Copilot `gemini-3-8-flash` **Google** · 2026-09-03 · frontier · reasoning _Flash-tier agent model enters Copilot under introductory provider pricing through December 31_ Treat current economics as a promotional observation, not a 2027 run rate. GitHub describes recovery from actionable terminal failures but publishes no reproducible benchmark or exact multiplier, so buyers should measure completed-task cost on their own harness. Source: [GitHub](https://github.blog/changelog/2026-09-03-gemini-3-8-flash-is-now-available-in-github-copilot/). ## Open weights ### Muse-Glimmer-30B NVFP4 `muse-glimmer-30b` **NVIDIA** · 2026-09-01 · open_frontier · multimodal _Apache-2.0 Blackwell checkpoint uses NVIDIA's 4-bit floating-point format (NVFP4) to cut storage while preserving multimodal tools_ This is a deployment-footprint update to an existing W33 model, not a new frontier row. Blackwell operators should test it for lower memory pressure, but NVIDIA's small Terminal-Bench gain over the 16-bit floating-point (BF16) baseline may be sampling noise and should not drive a capability claim. Source: [NVIDIA on Hugging Face](https://huggingface.co/nvidia/Muse-Glimmer-30B-NVFP4). ### NeoMME-800M **H Company** · 2026-09-03 · specialist · multimodal _Apache-2.0 single-tower multimodal encoder for shared text-token and raw-image-patch retrieval_ Relevant for multimodal retrieval and agent memory, not generative-model substitution. Day-zero Transformers support makes it testable now, but no production latency, retrieval cost, or independent benchmark is public, so keep it out of production scorecards. Source: [Hugging Face, NeoMME paper](https://huggingface.co/blog/Hcompany/neomme). ## Architecture watch ### Runtime version becomes a governed SKU property _Examples:_ GPT-6 Astra Standard harness, GPT-6 Astra Provider Adapter. ARC Prize's 35.9-point same-effort spread is too large to leave the runtime as an implementation footnote. Evaluation records should version context compression, memory visibility, state persistence, tool surface, retry budget, entitlement, and retention route with the model. Source: [ARC Prize, OpenAI](https://arcprize.org/blog/astra). ### Distribution policy is part of the model SKU _Examples:_ Claude Fable 5.1 administrator enablement, Gemini 3.8 Flash promotional pricing, Astra enterprise access off by default. Three closed-model movements arrived with material entitlement, retention, or price-window conditions. Procurement catalogs should stop recording only model family and API rate; they need surface, administrator state, retention route, promotion expiry, and provider adapter version. Source: [GitHub, OpenAI](https://github.blog/changelog/2026-09-01-claude-fable-5-1-generally-available-in-github-copilot/). ### Open-weight progress shifted from new lineage to deployment compression _Examples:_ Muse-Glimmer-30B NVFP4, NeoMME-260M, NeoMME-800M. W36 did not repeat W35's open-frontier release wave. NVIDIA cut an existing checkpoint's footprint 2.4× and H Company opened compact multimodal encoders; both lower implementation friction, but neither supplies an independent frontier comparison. Buyers should value deployability without relabeling it capability convergence. Source: [NVIDIA on Hugging Face, H Company](https://huggingface.co/nvidia/Muse-Glimmer-30B-NVFP4). ## Benchmark moves ### ARC-AGI-3 Semi-Private At matched maximum reasoning, Astra moved from 62.7% on the Standard harness to 98.6% on OpenAI's Provider Adapter, a 35.9-point system spread. - GPT-6 Astra, Provider Adapter maximum reasoning: 98.6% / $17,332 - GPT-6 Astra, Standard harness maximum reasoning: 62.7% / $26,098 Source: [ARC Prize](https://arcprize.org/blog/astra). ### Artificial Analysis Intelligence Index v4.1.1 Astra entered at 61 at maximum effort, reinforcing that one interactive benchmark should not define general capability. - Claude Fable 5.1: 66 - Claude Opus 5: 63 - GPT-6 Astra, maximum effort: 61 Source: [Artificial Analysis](https://artificialanalysis.ai/models/gpt-6-astra). ### Terminal-Bench 2.1, vendor-reported checkpoint comparison NVIDIA reports 47.05% for Muse-Glimmer NVFP4 versus 45.22% BF16, while cautioning that the small gain may be sampling noise. - Muse-Glimmer-30B NVFP4 (vendor-reported): 47.05% - Muse-Glimmer-30B BF16 (vendor-reported): 45.22% Source: [NVIDIA on Hugging Face](https://huggingface.co/nvidia/Muse-Glimmer-30B-NVFP4). ## Tier scorecard _As of 2026-09-05._ | Tier | Leader | Challenger | Note | |---|---|---|---| | Closed frontier | Claude Fable 5.1 | Claude Opus 5 | Artificial Analysis Index remains the broad capability anchor; Astra's adapter result is harness-specific and should not replace it. | | Open frontier | GLM-5.3-Flash | IBM Granite 4.2 30B | No new open-frontier release displaced W35's leaders; Muse-Glimmer NVFP4 changes footprint, not lineage or independent rank. | | Reasoning | Claude Fable 5.1 | GPT-6 Astra | Astra is the system-design challenger; persistent state materially changes ARC results, while Fable retains the broader independent Index lead. | | Coding | Claude Fable 5.1 | Gemini 3.8 Flash | Both gained GitHub distribution; compare completed-task cost after administrator policy and year-end Gemini pricing are included. | | Multimodal | Gemini 3.8 Flash | GPT-6 Astra | Astra adds 1M context and computer use; Gemini's Copilot rollout broadens surface reach, but GitHub publishes no neutral task benchmark. | | Edge / small | NeoMME-800M | NeoMME-260M | Specialist retrieval encoders, not generative leaders; test for multimodal memory only after measuring latency and retrieval quality. | ## Vendor signals - **2026-09-03 · OpenAI — Astra enterprise access off by default; Standard API priced at $10/$50 per million input/output tokens** - Astra is a premium governed rollout, not an automatic fleet replacement. Require explicit enablement, provider-adapter versioning, reasoning-effort capture, and a portability test before using its adapter result in procurement. - _Source:_ [OpenAI](https://openai.com/index/gpt-6-astra/) - **2026-09-01 · Anthropic / GitHub — Fable 5.1 reaches general availability but Business and Enterprise administrators must enable it** - Public availability does not equal enterprise entitlement. Catalog the selected safety and retention route, including the documented 30-day default outside approved zero-data-retention arrangements. - _Source:_ [GitHub](https://github.blog/changelog/2026-09-01-claude-fable-5-1-generally-available-in-github-copilot/) - **2026-09-03 · Google / GitHub — Gemini 3.8 Flash enters Copilot under introductory provider pricing through December 31** - Do not annualize current economics into 2027. Measure completed-task cost now, then rerun the same workload when post-promotion pricing is published. - _Source:_ [GitHub](https://github.blog/changelog/2026-09-03-gemini-3-8-flash-is-now-available-in-github-copilot/) - **2026-09-01 · NVIDIA — Muse-Glimmer-30B NVFP4 ships Apache 2.0 for Blackwell with a 24.7 GB checkpoint** - Blackwell self-hosters gain a smaller deployable artifact. Treat the vendor-reported benchmark delta as noise until reproduced, and value the release on footprint and runtime compatibility. - _Source:_ [NVIDIA on Hugging Face](https://huggingface.co/nvidia/Muse-Glimmer-30B-NVFP4) ## Watchlist - **Sep 9 — GLM-5.3-Flash promotion expires.** Recalculate the W35 open-frontier cost case on steady-state list pricing and independent provider variance. - **September — Neutral Astra memory-harness reproductions.** A visible-state harness closing part of the 35.9-point matched-effort gap would make the runtime gain more portable. - **Q4 2026 — Anthropic Enterprise Frontier Safeguards rollout.** Look for precision, recall, rolling-window, alert-volume, audit, and customer-cloud cost evidence before approval. - **Dec 31 — Gemini 3.8 Flash introductory pricing expiry.** The post-promotion rate determines whether today's Copilot task economics survive into 2027. - **Next AA refresh — Astra and W35 open-model independent rankings.** Use a common harness to test whether Astra's system advantage and W35's vendor-reported open scores generalize. ## Changelog - Added GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash to the LLM tree. - Held the open-frontier scorecard after a deployment-compression week; Muse-Glimmer NVFP4 remains an update to the W33 lineage. - Credited ARC Prize for the harness finding and reframed the issue around SKU governance: entitlement, retention route, runtime version, effort, and price window. --- Source of truth: `src/data/industry/models/2026-W36.ts`. Canonical HTML: . PDF: . Tree: .