--- title: >- The serving layer changed underneath models already deployed, and this week's new checkpoints lead on cost and footprint rather than capability publication: The Model Pulse slug: 2026-W37 issueNumber: 21 isoYear: 2026 isoWeek: 37 cadence: weekly publishedAt: '2026-09-12' periodLabel: Week 37 of 2026 canonicalUrl: https://brianletort.ai/industry/models/2026-W37 pdfUrl: https://brianletort.ai/downloads/model-pulse-2026-W37.pdf schemaVersion: 2026.05.02 treeDelta: added: - deepseek-v4-1-flash - nex-n2-5-max - nex-n2-5-mini updated: - deepseek-v4-pro - deepseek-v4-flash frontierMovements: - modelId: null name: GPT-Live-1 vendor: OpenAI releaseDate: '2026-09-10' tier: frontier architecture: multimodal - modelId: gpt-6-astra name: GPT-6 Astra vendor: OpenAI releaseDate: '2026-09-03' tier: frontier architecture: reasoning openWeights: - modelId: deepseek-v4-1-flash name: DeepSeek-V4.1-Flash vendor: DeepSeek releaseDate: '2026-09-10' tier: open_frontier architecture: moe - modelId: nex-n2-5-max name: Nex-N2.5 Max vendor: Nex-AGI releaseDate: '2026-09-08' tier: open_frontier architecture: moe architecturePatterns: - Cache residency, not parameter count, is the long-context lever - The conversational layer separates from the reasoning layer - The serving floor decides who can actually use an open model - Inference silicon routes around high-bandwidth memory rather than competing for it benchmarks: - Artificial Analysis Intelligence Index v4.3 - Cost per Intelligence Index task (v4.3) - Terminal-Bench v4.0 - DeepSWE v1.1 (vendor-reported) scorecardAsOf: '2026-09-12' vendorSignals: - vendor: DeepSeek date: '2026-09-10' signal: Flagship endpoint rerouted to a smaller model, and its predecessor retired the same day - vendor: GitHub date: '2026-09-10' signal: A model deprecated the same day across chat, inline edits, ask and agent modes - vendor: Artificial Analysis date: '2026-09-07' signal: Second Intelligence Index revision in four days, with private-test weighting raised to 45% - vendor: OpenAI date: '2026-09-10' signal: Agents API ships in public beta with US-only residency and no zero data retention - vendor: Anthropic date: '2026-09-09' signal: Lab discloses that its own agentic transcript review missed evaluation incidents - vendor: European Commission date: '2026-09-10' signal: >- EU cybersecurity agency granted hands-on access to two frontier models for vulnerability testing --- # The serving layer changed underneath models already deployed, and this week's new checkpoints lead on cost and footprint rather than capability *The Model Pulse · Issue 21 · Week 37 of 2026 · Published 2026-09-12* ## Big Read W37 produced three material changes to models already deployed and two genuinely new open-weight releases, neither of which leads on capability. DeepSeek retired V4-Flash on September 10 and announced that from 04:00 UTC on September 14 every deepseek-v4-pro request routes to V4.1-Flash and bills at V4.1-Flash rates until a V4.1-Pro exists. GitHub deprecated a model across chat, inline edits, ask and agent modes on September 10, with a replacement that administrators may have to enable themselves. Artificial Analysis revised its Intelligence Index on September 7, four days after the previous revision, rescaling the composite so that last week's 61 and this week's 53 are not the same measurement. Read together, those three are one event: the identity of the model behind a pinned endpoint is now a vendor-controlled variable, and so is the scale of the number you used to justify choosing it. None of the three required a new model to be released. All three change what an already-deployed system does. The genuine releases of the week reinforce the point rather than contradicting it. DeepSeek published an MIT-licensed 552B multimodal Mixture-of-Experts model whose headline property is a key-value cache of 890 bytes per token, roughly a quarter of its predecessor's, which is a serving-economics claim rather than a capability claim. OpenAI made a full-duplex voice model generally available at $0.05 per minute and split the conversational layer from the reasoning layer, so a voice agent is now two rate cards that can change independently. Nex-AGI published a 1.6 trillion-parameter open checkpoint whose own reference deployment requires sixteen H200-class accelerators, which makes it nominally open and practically gated by hardware. W36 already argued that model diligence had moved off the model card and onto the harness, so the direction is not new. What W37 adds is the other three surfaces and a date for each: an endpoint reassigned to a different model on September 14, a managed service whose residency and retention posture is documented rather than inferred, and an index whose version number now changes the score. The defensible W37 read is that harness diligence has become endpoint, residency and index-version diligence as well, and that all four moved inside one week. ## Tree delta Three rows added and two updated. The additions are the week's two open-weight release events; the updates are both in the DeepSeek V4 line and are the more consequential half of the delta, because they change what an already-deployed endpoint serves. **Added (3):** `deepseek-v4-1-flash`, `nex-n2-5-max`, `nex-n2-5-mini`. **Updated (2):** `deepseek-v4-pro`, `deepseek-v4-flash`. *deepseek-v4-1-flash enters as an MIT-licensed 552B multimodal Mixture-of-Experts at 1M context, and nex-n2-5-max and nex-n2-5-mini enter as the two Nex-N2.5 checkpoints whose weights were actually published. The 397B Nex-N2.5 Pro is deliberately not added: its weights were listed as coming soon inside the window and it was reachable only through hosted access, so there is nothing yet to place. On the update side, deepseek-v4-flash is retired as of September 10 and deepseek-v4-pro records that its endpoint routes to a different, smaller model from September 14 with no change required by the caller. Read the three additions as cost and footprint claims rather than capability claims, because that is what both releases lead with.* ## Frontier movements ### GPT-Live-1 **OpenAI** · 2026-09-10 · frontier · multimodal _Full-duplex voice generally available at $0.05 per minute, with reasoning delegated to a separately billed model_ The model listens and speaks simultaneously and hands reasoning and tool calls to a backend text model the developer selects and pays for separately. That makes a voice agent a two-line bill whose latency and cost can be tuned independently, and it also means the deployment now depends on two rate cards that can change without notice. OpenAI reports a 30-percentage-point improvement on a full-duplex benchmark over its predecessor, which is vendor-reported. Source: [OpenAI](https://openai.com/index/introducing-gpt-live-1-in-the-api/). ### GPT-6 Astra `gpt-6-astra` **OpenAI** · 2026-09-03 · frontier · reasoning _Re-measured mid-window at an identical rounded score to Claude Fable 5.1 but 57% lower measured cost per task_ Artificial Analysis Intelligence Index v4.3, published September 7, puts Astra and Fable 5.1 level at 53 while measuring cost per index task at $3.26 against $7.63. The rank is an artifact of a benchmark revised twice in four days and should not be read as model progress. The cost separation is the durable procurement finding, and it survives the rescaling because it is a measured spend figure rather than a composite score. Source: [Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3). ## Open weights ### DeepSeek-V4.1-Flash `deepseek-v4-1-flash` **DeepSeek** · 2026-09-10 · open_frontier · moe _MIT-licensed 552B multimodal Mixture-of-Experts with a 1M context and an 890-byte-per-token cache_ The defining number is not a benchmark but a memory figure: DeepSeek reports a global key-value cache of 890 bytes per token, roughly a quarter of its predecessor's, which is what makes a million-token context economically serveable rather than merely supported. It exceeds the larger V4-Pro it replaces on two of DeepSeek's own reported benchmarks, DeepSWE v1.1 at 74.2 resolved against 62.7 and Terminal-Bench 2.1 at 90.6 Pass@1 against 87.9, both vendor-run and neither independently reproduced. The weights are published under MIT, so the serving economics are available to anyone with the hardware. Note a disclosed inconsistency: the model card states a 552B backbone while repository metadata lists 763B, and DeepSeek does not explain the difference. Source: [DeepSeek model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash). ### Nex-N2.5 Max `nex-n2-5-max` **Nex-AGI** · 2026-09-08 · open_frontier · moe _1.6 trillion-parameter open checkpoint whose own reference deployment needs sixteen H200-class accelerators_ This is the clearest current example of an open weight that is nominally available and practically gated: the vendor's reference serving configuration specifies two nodes of eight H200 GPUs with tensor and expert parallelism at sixteen. Read it as a post-training achievement rather than a new foundation, because independent coverage reports it is built on DeepSeek-V4-Pro-Base, and its 1.6 trillion total and 49 billion active parameters match that model exactly. For an organisation that already holds multi-node H200 capacity it is a real self-hosting option at trillion-parameter scale. For everyone else the 35B sibling is the only member of the family that can actually be piloted. Source: [Nex-AGI model card](https://huggingface.co/nex-agi/Nex-N2.5-Max). ## Architecture watch ### Cache residency, not parameter count, is the long-context lever _Examples:_ DeepSeek-V4.1-Flash, GLM 5.3 under vLLM Hybrid HiSparse. Two independent pieces of work in the same week attacked long context through memory policy rather than model size. DeepSeek reports a global key-value cache of 890 bytes per token, roughly a quarter of its predecessor's, at a 1,048,576-token context. vLLM published a residency policy that keeps cache on the accelerator while capacity allows and offloads all but indexer-selected tokens to host memory under pressure, which it states lets GLM 5.3 run its full million-token context on a single eight-GPU node where that was previously impossible. Both reframe long context as a bytes-per-token budgeting problem, which is the number to put in a capacity plan. Source: [vLLM, DeepSeek model card](https://vllm.ai/blog/2026-09-08-glm53-part1-hybrid-sparse-offloading). ### The conversational layer separates from the reasoning layer _Examples:_ GPT-Live-1. A full-duplex voice model priced per minute, delegating reasoning and tool calls to a separately selected and separately billed text model, splits what used to be one procurement into two. The upside is real: cheap models can handle scheduling turns while expensive ones handle escalations, and latency can be tuned where it is felt. The cost is that the deployment's economics now sit on two rate cards, either of which can be repriced or rerouted by its vendor, and this week supplied an example of exactly that happening elsewhere in the stack. Source: [OpenAI](https://openai.com/index/introducing-gpt-live-1-in-the-api/). ### The serving floor decides who can actually use an open model _Examples:_ Nex-N2.5 Max, Nex-N2.5 mini, DeepSeek-V4.1-Flash. Open weights published this week span a two-order-of-magnitude range in what it takes to run them, and the licence tells you almost nothing about that. A 1.6 trillion-parameter checkpoint arrives with a sixteen-accelerator reference configuration, a 35B sibling arrives as the pilotable member of the same family, and a 552B model arrives with a cache figure specifically low enough to make its full context affordable. The practical reading is that openness and accessibility have decoupled, so an open-weight strategy now needs a hardware floor stated alongside every candidate model. Source: [Nex-AGI model card, DeepSeek model card](https://huggingface.co/nex-agi/Nex-N2.5-mini). ### Inference silicon routes around high-bandwidth memory rather than competing for it _Examples:_ d-Matrix Raptor, Positron Asimov. Two inference-silicon events in the same week share one design thesis: avoid the constrained memory and advanced-packaging supply chain instead of bidding inside it. One uses logic on three-dimensional DRAM and enters through an established rack interconnect ecosystem, with a tape-out targeted before the end of 2026 and rack availability in the fourth quarter of 2027. The other uses standard low-power memory to reach 288 GB to 2,304 GB per chip, taping out at the end of 2026 for production in the second half of 2027. Neither has silicon; both are bets that memory supply rather than compute is the binding constraint on cost per token. Source: [d-Matrix, Positron AI](https://www.d-matrix.ai/announcements/d-matrix-rackscale-nvidia/). ## Benchmark moves ### Artificial Analysis Intelligence Index v4.3 Rescaled twice in four days, so last week's 61 and this week's 53 are different measurements and any reading of a decline is wrong - Claude Fable 5.1 (max, fallback): 53 - GPT-6 Astra (max): 53 - Claude Opus 5 (max): 51 - Muse Spark 1.3 (max): 48 - GLM-5.3 Flash (open weights): 42 Source: [Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3). ### Cost per Intelligence Index task (v4.3) The durable finding of the week is price rather than rank: a 57% separation at an identical rounded score - GPT-6 Astra (max): $3.26 - Claude Fable 5.1 (max, fallback): $7.63 Source: [Artificial Analysis](https://artificialanalysis.ai/evaluations/artificial-analysis-intelligence-index). ### Terminal-Bench v4.0 Upgraded from v2.1 inside the same index revision, so prior Terminal-Bench figures do not carry forward into comparisons - GPT-6 Astra (max): 59.1% - Claude Fable 5.1: 52.0% - Claude Opus 5 (max): 49.0% Source: [Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3). ### DeepSWE v1.1 (vendor-reported) A smaller successor beats the larger flagship it replaces on the vendor's own harness, which is the stated basis for retiring the flagship endpoint - DeepSeek-V4.1-Flash (max effort): 74.2 resolved - DeepSeek-V4-Pro: 62.7 resolved Source: [DeepSeek model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash). ## Tier scorecard _As of 2026-09-12._ | Tier | Leader | Challenger | Note | |---|---|---|---| | Closed frontier | Claude Fable 5.1 | GPT-6 Astra | Level at 53 on Intelligence Index v4.3, but the tie is a rounding artifact on a scale revised twice in four days; treat the $3.26 against $7.63 cost-per-task spread as the real separation. | | Open frontier | DeepSeek-V4.1-Flash | Nex-N2.5 Max | MIT weights at 552B with an 890-byte-per-token cache against a 1.6T checkpoint needing sixteen H200-class accelerators; the licence is no longer the constraint, the serving floor is. Independent coverage of Nex-AGI's materials, not a Nex-AGI page, reports that the challenger is post-trained from DeepSeek's own V4-Pro base; its parameter shape matches exactly, which makes this tier closer to one lineage at two price points than to two competitors. | | Reasoning | Claude Fable 5.1 | GPT-6 Astra | No reasoning-specific release landed in the window, so the standing order is unchanged and rests on an index whose composition changed underneath it on September 7. | | Coding | GPT-6 Astra | Claude Fable 5.1 | Astra leads Terminal-Bench v4.0 at 59.1% against 52.0%, on a benchmark version introduced this week, so this ordering has no comparable prior-week figure. | | Multimodal | Gemini 3.8 Flash | DeepSeek-V4.1-Flash | Native video input remains the Flash family's distinguishing modality rather than a single model's, and DeepSeek's new multimodal open weight is the first credible open alternative at frontier scale. | | Edge / small | Nex-N2.5 mini | Desert Ant on-device model set | A 35B open checkpoint against eighteen task-specific models that run entirely on device with no per-inference bill below a stated free tier; both are pilotable without data centre capacity. | ## Vendor signals - **2026-09-10 · DeepSeek — Flagship endpoint rerouted to a smaller model, and its predecessor retired the same day** - From 04:00 UTC on September 14 every deepseek-v4-pro request serves V4.1-Flash at V4.1-Flash rates until a V4.1-Pro ships. Any team with the flagship pinned in code gets a different model and a different bill without changing a line, which makes endpoint identity a term to pin contractually rather than a constant. - _Source:_ [DeepSeek API change log](https://api-docs.deepseek.com/updates) - **2026-09-10 · GitHub — A model deprecated the same day across chat, inline edits, ask and agent modes** - The replacement may need enterprise administrator enablement, so an organisation running scheduled or unattended agent work can have that work change behaviour without shipping a release of its own. Version the model your automations depend on and test scheduled agents against model changes you do not control. - _Source:_ [GitHub Changelog](https://github.blog/changelog/label/copilot/) - **2026-09-07 · Artificial Analysis — Second Intelligence Index revision in four days, with private-test weighting raised to 45%** - The index replaced a banking benchmark with a 657-task held-out business-workflow set and upgraded Terminal-Bench two major versions. Any model-selection business case quoting an Artificial Analysis score from before September 7 must now carry its index version or it is not defensible in review. - _Source:_ [Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3) - **2026-09-10 · OpenAI — Agents API ships in public beta with US-only residency and no zero data retention** - OpenAI's own documentation states the managed harness supports US data residency only and does not support zero data retention, and that choosing a self-hosted sandbox does not make it eligible. That disqualifies it for regulated workloads today regardless of capability, which makes this a developer-velocity decision rather than a production one. - _Source:_ [OpenAI Developer Docs](https://developers.openai.com/api/docs/guides/agents-api/overview) - **2026-09-09 · Anthropic — Lab discloses that its own agentic transcript review missed evaluation incidents** - Anthropic states its original review spanned roughly 141,000 transcripts and that because the scan relied on agentic search it missed a set of transcripts that also had internet access, surfacing a fourth incident only while assembling material for an external evaluator. It is a direct argument for deterministic coverage checks on any agentic evidence review used in audit. - _Source:_ [Anthropic](https://www.anthropic.com/news/alignment-assessment-cybersecurity-incidents) - **2026-09-10 · European Commission — EU cybersecurity agency granted hands-on access to two frontier models for vulnerability testing** - Access followed one model's release by about a week and the other by more than five months from announcement, with roughly three months of safeguard negotiation. The gap shows the arrangement is negotiated bilaterally rather than compelled, so regulator visibility currently scales with each vendor's willingness rather than a uniform standard. - _Source:_ [Reuters](https://www.reuters.com/technology/eus-cybersecurity-agency-granted-access-mythos-5-ai-model-commission-says-2026-09-10/) ## Watchlist - **Sep 14 — DeepSeek flagship reroute takes effect.** Every deepseek-v4-pro request begins serving V4.1-Flash at V4.1-Flash rates; re-baseline any evaluation or cost model pinned to the flagship endpoint. - **Sep 15-17 — AI Infra Summit silicon status.** Marvell's 102 Tbps Ethernet switch and 256-lane scale-up switch were announced with no silicon status, node, power, or availability date; the show is where those should appear. - **Sep 21-23 — First multi-vendor 1600ZR interoperability demonstrations.** The Implementation Agreement published September 9 has no shipping module against it yet, so interop results are the first real test of whether 1.6T coherent is procurable multi-vendor. - **Later in 2026 — vLLM v0.30 and the cache-residency policy.** Hybrid HiSparse is not in any release and needs a pinned checkout today, so the single-node million-token capacity claim only enters a plan when v0.30 ships. - **Dec 31, 2026 — Gemini 3.8 Flash introductory pricing expiry.** Current Flash-tier task economics rest on introductory provider pricing; the defensible 2027 run rate is whatever rate survives the reversal. ## Changelog - Sourced from the W37 graded event set and the counterbrief, with an evidence cutoff of September 12 and a September 7 to 12 observation window. - Three tree rows added and two updated. The delta records release events and endpoint changes; it is not used anywhere in this issue as evidence about release cadence or capability progress, because a maintained artifact's row count measures maintenance state rather than the industry. - The Nex-N2.5 family is carried as two rows, not three. Nex-N2.5 Pro is announced at 397B but its weights were listed as coming soon inside the window and it was reachable only through hosted access, so it is recorded nowhere rather than given a status it has not earned. - Both Nex rows are recorded as post-trained derivatives rather than new foundations, Max from DeepSeek-V4-Pro-Base and mini from a Qwen3.5 variant, on independent coverage opened directly. That reframes the open-frontier tier: the challenger runs on the leader's own base model, and the scorecard note says so. - Intelligence Index figures now carry their index version inline. Following two revisions in four days, v4.1.1 scores quoted in W36 are not comparable to the v4.3 scores quoted here, and the issue states that explicitly rather than showing a trend line. - Review correction, continuity. The closing read previously presented the move of model diligence off the model card as a W37 finding. W36 had already made that call about the harness. The issue now credits it and claims only what W37 adds, which is the endpoint, the residency policy and the index version, each with a date. The DeepSeek performance claim also now names the two vendor-run benchmarks behind it, DeepSWE v1.1 and Terminal-Bench 2.1, with both figures, rather than asserting that it exceeds its predecessor on an unnamed harness. - Endpoint identity is added to the diligence list alongside model, harness, memory policy, and retention, on the basis of the DeepSeek reroute and the same-day model deprecation inside agent mode. --- Source of truth: `src/data/industry/models/2026-W37.ts`. Canonical HTML: . PDF: . Tree: .