{
  "_meta": {
    "publication": "The Model Pulse",
    "schemaVersion": "2026.05.02",
    "generatedAt": "2026-09-08T22:45:10.948Z",
    "canonicalUrl": "https://brianletort.ai/industry/models/2026-W24",
    "markdownUrl": "https://brianletort.ai/industry/models/2026-W24/llm.md",
    "pdfUrl": "https://brianletort.ai/downloads/model-pulse-2026-W24.pdf",
    "treeUrl": "https://brianletort.ai/industry/tree",
    "sourceFile": "src/data/industry/models/2026-W24.ts"
  },
  "issue": {
    "slug": "2026-W24",
    "isoYear": 2026,
    "isoWeek": 24,
    "issueNumber": 8,
    "publishedAt": "2026-06-13",
    "cadence": "weekly",
    "periodLabel": "Week 24 of 2026",
    "bigRead": {
      "headline": "A new closed frontier shipped and was switched off in the same week; the durable open progress was a diffusion efficiency path.",
      "body": "W24's model story had two halves. Anthropic shipped Claude Fable 5 — a new Mythos-class tier above Opus — and it debuted #1 on the independent Artificial Analysis Intelligence Index at 64.9, roughly five points ahead of GPT-5.5. Three days later (Jun 12), a U.S. government export-control directive forced Anthropic to disable Fable 5 and its safeguards-lifted Mythos 5 sibling for every customer, routing queries back to Opus 4.8. So the public closed-frontier leader for half a week is, at week's end, unavailable — the working default is again Opus 4.8. Meanwhile the open layer's real progress was efficiency, not a new crown: Google DeepMind released DiffusionGemma, an Apache-2.0 text-diffusion model that generates ~1,000+ tok/s on an H100, and Cohere shipped North Mini Code, a cheap self-host coding MoE; MiniMax-M3 was announced (AA Index 55) but its weights are still pending. Gemini 3.5 Pro remained not-GA. The procurement implication is sharper than another leaderboard reshuffle: a top-tier closed model can now be revoked by a third party, so production systems need a standardized eval harness and a hard fallback router, with cheap open/local models carrying routine work and the closed frontier reserved for high-value reasoning."
    },
    "treeDelta": {
      "summary": "Three W24 additions: Claude Fable 5 (new Mythos-class closed frontier), DiffusionGemma (open text-diffusion), and Cohere North Mini Code (efficient open coding MoE).",
      "added": [
        "claude-fable-5",
        "diffusiongemma",
        "north-mini-code"
      ],
      "updated": [],
      "note": "Mythos 5 is the same model as Fable 5 with cyber safeguards lifted (government-only) and is not a separate tree row. MiniMax-M3 is excluded until its weights are released, and Gemini 3.5 Pro is excluded until Google publishes a GA model card or API identifier."
    },
    "frontierMovements": [
      {
        "name": "Claude Fable 5",
        "vendor": "Anthropic",
        "releaseDate": "2026-06-09",
        "headline": "New Mythos-class tier above Opus; debuted #1 on the Artificial Analysis Intelligence Index at 64.9, priced $10/$50 per MTok",
        "why": "Fable 5 is the new public top of the closed frontier on independent indices, but it is priced as an async heavy-lift tool (~2x Opus 4.8) and was free on paid tiers only through Jun 22 before reverting to usage credits. Architects should reserve it for high-value reasoning and codebase-scale orchestration, not default every agent call to it, and should re-benchmark on cost-per-task rather than headline scores.",
        "tier": "frontier",
        "architecture": "reasoning",
        "source": "Anthropic launch; Artificial Analysis Intelligence Index",
        "sourceUrl": "https://www.anthropic.com/news/claude-fable-5-mythos-5"
      },
      {
        "name": "Claude Mythos 5 (and the Jun 12 takedown)",
        "vendor": "Anthropic",
        "releaseDate": "2026-06-12",
        "headline": "Same model as Fable 5 with safeguards lifted for a government program; both force-disabled globally by a U.S. export-control directive three days after launch",
        "why": "This is the first known government-forced takedown of a deployed frontier model, and it reframes model selection: a top-tier closed model is now a sovereign/regulatory single point of failure that a third party can switch off. Architects and operators should build hard fallbacks (queries already revert to Opus 4.8) and avoid single-sourcing the frontier for any production-critical path.",
        "tier": "frontier",
        "architecture": "reasoning",
        "source": "VentureBeat; MarkTechPost (Reuters/BBC/Axios cited secondhand)",
        "sourceUrl": "https://venturebeat.com/technology/anthropic-blocks-all-public-access-to-claude-fable-5-mythos-5-following-us-government-order-what-enterprises-should-do"
      },
      {
        "name": "Gemini 3.5 Pro",
        "vendor": "Google DeepMind",
        "releaseDate": "2026-06 target",
        "headline": "Still pending at W24 close: limited Vertex enterprise preview, no public API identifier, price row, or independent benchmark despite a Google-stated June target",
        "why": "A frontier movement by absence for the second consecutive issue. Buyers should keep the launch on the June watchlist but not pause current baselines on an unreleased SKU; the likely enterprise pattern remains task routing (Gemini for huge-context/multimodal, Anthropic/OpenAI for coding and agentic reliability) rather than a universal replacement.",
        "tier": "frontier",
        "architecture": "reasoning",
        "source": "TechTimes; Presenc launch brief",
        "sourceUrl": "https://www.techtimes.com/articles/317919/20260606/google-gemini-35-pro-nears-june-launch-2-million-token-context-deep-think-reasoning.htm"
      }
    ],
    "openWeights": [
      {
        "modelId": "diffusiongemma",
        "name": "DiffusionGemma",
        "vendor": "Google DeepMind",
        "releaseDate": "2026-06-10",
        "headline": "Apache-2.0 open text-diffusion model: 26B / ~3.8B-active MoE on a Gemma 4 backbone, multimodal-in, 256K context, ~1,000+ tok/s on an H100 with native vLLM",
        "why": "DiffusionGemma is the week's most architecturally significant open release: a credible open path to low-latency, local interactive inference via block diffusion rather than autoregressive decoding. Teams building latency-sensitive local agents should evaluate it as an efficiency option and watch whether the autoregressive-vs-diffusion split becomes a real procurement axis.",
        "tier": "specialist",
        "architecture": "moe",
        "source": "Google blog; Hugging Face model card",
        "sourceUrl": "https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/"
      },
      {
        "modelId": "north-mini-code",
        "name": "North Mini Code",
        "vendor": "Cohere",
        "releaseDate": "2026-06-09",
        "headline": "Small open-weight coding MoE (~30B / ~3B active) scoring 27.6 on the Artificial Analysis Intelligence Index",
        "why": "North Mini Code is not frontier-competitive, but it is a cheap self-host candidate for routine code and text sub-agent calls. Teams trying to cut orchestration cost should treat it as a routing node for low-risk repeated work, reserving closed flagships for high-stakes reasoning.",
        "tier": "edge",
        "architecture": "moe",
        "source": "Artificial Analysis",
        "sourceUrl": "https://artificialanalysis.ai/articles/claude-fable-5-mythos"
      },
      {
        "name": "MiniMax-M3 (announced; weights pending)",
        "vendor": "MiniMax",
        "releaseDate": "2026-06-08",
        "headline": "Announced at 55 on the Artificial Analysis Intelligence Index — would be the leading open-weights model once weights release, but is not yet downloadable",
        "why": "MiniMax-M3 is the strongest competitively-scoring open item of the window, but with weights pending it is a signal, not a deployable asset. Procurement teams tracking an open-frontier alternative should prepare an evaluation slot but not plan a migration until the weights and an independent benchmark land.",
        "tier": "open_frontier",
        "architecture": "moe",
        "source": "Artificial Analysis",
        "sourceUrl": "https://artificialanalysis.ai/articles/claude-fable-5-mythos"
      }
    ],
    "architectureWatch": [
      {
        "pattern": "Government as a model kill switch",
        "examples": [
          "Claude Fable 5 takedown",
          "Claude Mythos 5 (Project Glasswing)"
        ],
        "body": "The Jun 12 export-control directive that forced Anthropic to disable Fable 5 and Mythos 5 globally turns regulatory access into an architecture variable. A model that is best-in-class on Monday can be unavailable by Friday for reasons outside the vendor's control. Enterprise architecture should now treat top-tier closed models as revocable dependencies and design fallback routing and capability-degradation plans accordingly.",
        "source": "VentureBeat; MarkTechPost"
      },
      {
        "pattern": "Text diffusion enters open weights",
        "examples": [
          "DiffusionGemma",
          "Gemma 4 backbone"
        ],
        "body": "DiffusionGemma brings block-diffusion text generation into the open-weight mainstream with native vLLM support and ~1,000+ tok/s throughput. The pattern matters because it offers a different latency/throughput profile than autoregressive decoding for interactive and local workloads. Architects evaluating local inference should add a diffusion option to their benchmark set rather than assuming autoregressive is the only path.",
        "source": "Google blog; Hugging Face"
      },
      {
        "pattern": "Frontier shipping concentrated in one vendor",
        "examples": [
          "Anthropic Fable 5",
          "Gemini 3.5 Pro pending",
          "GPT-5.6 rumor only"
        ],
        "body": "For the second straight issue, the only new frontier model came from Anthropic, while Google's Pro slipped and OpenAI's next model stayed a rumor. That concentration means the frontier's cadence — and now its availability — depends heavily on a single lab. Buyers should diversify model routing across vendors and keep a credible open/local tier so a single vendor's release or takedown cannot stall production.",
        "source": "Anthropic launch; Artificial Analysis; TechTimes"
      }
    ],
    "benchmarkMoves": [
      {
        "benchmark": "Artificial Analysis Intelligence Index",
        "movement": "Claude Fable 5 debuted #1 at 64.9, roughly five points ahead of GPT-5.5, and set the highest score on five of the ten underlying benchmarks",
        "rows": [
          {
            "model": "Claude Fable 5",
            "score": "64.9"
          },
          {
            "model": "GPT-5.5",
            "score": "~60"
          },
          {
            "model": "Gemini 3.1 Pro",
            "score": "~57"
          }
        ],
        "source": "Artificial Analysis",
        "sourceUrl": "https://artificialanalysis.ai/articles/claude-fable-5-mythos-intelligence-index"
      },
      {
        "benchmark": "SWE-Bench Pro (coding)",
        "movement": "Fable 5's vendor-scaffold 80.3% widens the headline closed lead, but it is unreplicated by Scale SEAL/Epoch; independent vals.ai shows 95% on SWE-Bench Verified",
        "rows": [
          {
            "model": "Claude Fable 5 (vendor scaffold)",
            "score": "80.3%"
          },
          {
            "model": "Claude Opus 4.8",
            "score": "69.2%"
          },
          {
            "model": "GPT-5.5",
            "score": "58.6%"
          }
        ],
        "source": "morphllm / vals.ai; vendor harness (contested)",
        "sourceUrl": "https://www.morphllm.com/swe-bench-pro"
      },
      {
        "benchmark": "Open-weight intelligence",
        "movement": "MiniMax-M3 (55, weights pending) would lead open weights once released; North Mini Code lands at 27.6 in the efficient tier",
        "rows": [
          {
            "model": "MiniMax-M3 (pending)",
            "score": "55"
          },
          {
            "model": "North Mini Code",
            "score": "27.6"
          }
        ],
        "source": "Artificial Analysis",
        "sourceUrl": "https://artificialanalysis.ai/articles/claude-fable-5-mythos"
      }
    ],
    "scorecard": {
      "asOf": "2026-06-13",
      "rows": [
        {
          "tier": "Closed frontier",
          "leader": "Claude Opus 4.8",
          "challenger": "Claude Fable 5 (suspended)",
          "note": "Fable 5 debuted #1 (AA 64.9) but was force-disabled Jun 12; Opus 4.8 is the working default again until access is restored."
        },
        {
          "tier": "Open frontier",
          "leader": "DeepSeek V4-Pro",
          "challenger": "MiniMax-M3 (weights pending)",
          "note": "No open frontier text model displaced the April leaders in-window; MiniMax-M3's 55 AA Index would contend once weights ship."
        },
        {
          "tier": "Reasoning",
          "leader": "Claude Opus 4.8",
          "challenger": "GPT-5.5",
          "note": "Fable 5 set the reasoning bar before its takedown; with it suspended, Opus 4.8 leads the available closed reasoning tier."
        },
        {
          "tier": "Coding",
          "leader": "Claude Opus 4.8",
          "challenger": "GPT-5.5",
          "note": "Fable 5's 80.3% SWE-Bench Pro is a contested vendor figure and the model is suspended; Opus 4.8 remains the dependable coding default."
        },
        {
          "tier": "Multimodal",
          "leader": "Gemini 3.1 Pro",
          "challenger": "DiffusionGemma",
          "note": "Gemini remains the general multimodal reference; DiffusionGemma opens a low-latency open diffusion branch."
        },
        {
          "tier": "Edge / small",
          "leader": "Mellum2",
          "challenger": "North Mini Code",
          "note": "Efficient open coding/sub-agent models keep widening; North Mini Code joins Mellum2 as a cheap self-host routing node."
        }
      ]
    },
    "vendorSignals": [
      {
        "vendor": "Anthropic",
        "date": "2026-06-12",
        "signal": "A U.S. export-control directive forced Anthropic to disable Fable 5 and Mythos 5 for all customers; queries fall back to Opus 4.8",
        "meaning": "Top-tier model availability is now a sovereign/regulatory variable, not just an uptime SLA. Buyers should require contractual clarity on access continuity and design fallback routing before standardizing on any single closed flagship.",
        "source": "VentureBeat; MarkTechPost",
        "sourceUrl": "https://venturebeat.com/technology/anthropic-blocks-all-public-access-to-claude-fable-5-mythos-5-following-us-government-order-what-enterprises-should-do"
      },
      {
        "vendor": "Anthropic",
        "date": "2026-06-09",
        "signal": "Fable 5 priced at $10/$50 per MTok (~2x Opus 4.8); included on paid tiers only through Jun 22, then reverting to usage credits",
        "meaning": "The top-tier price ceiling moved up and the free-access window is short. Operators should model the post-Jun 22 cost cliff and route only high-value work to Fable 5, keeping cheaper models for routine calls.",
        "source": "Anthropic launch",
        "sourceUrl": "https://www.anthropic.com/news/claude-fable-5-mythos-5"
      },
      {
        "vendor": "xAI",
        "date": "2026-06-11",
        "signal": "Launched the Grok Build Plugin Marketplace bundling skills, MCP servers, and agents (commit-SHA-pinned), with partners MongoDB, Vercel, Sentry, Chrome DevTools, and Cloudflare",
        "meaning": "xAI is competing on agent tooling and distribution rather than a new frontier model this window. Platform teams should weigh marketplace lock-in and pinning guarantees when evaluating agent ecosystems.",
        "source": "x.ai",
        "sourceUrl": "https://x.ai/news/grok-plugin-marketplace"
      },
      {
        "vendor": "OpenAI",
        "date": "2026-06-11",
        "signal": "Acquired Ona (Gitpod) for cloud sandboxes that keep long-running agents alive, folding it into Codex (5M+ weekly users)",
        "meaning": "Durable long-running-agent infrastructure is being absorbed into the model layer, narrowing 'agent persistence' as a standalone wedge. Architects building long-horizon agents should expect this capability to become a platform feature rather than a separate purchase.",
        "source": "SiliconANGLE",
        "sourceUrl": "https://siliconangle.com/2026/06/11/openai-acquires-ai-agent-orchestration-startup-ona/"
      }
    ],
    "watchlist": [
      {
        "window": "Jun 14-30",
        "title": "Fable 5 / Mythos 5 access restoration",
        "why": "Whether Anthropic restores a top-tier model — and with what geo-gating or KYC controls — sets the precedent for sovereign access risk and determines whether Fable 5 re-enters the closed-frontier scorecard."
      },
      {
        "window": "Jun 14-30",
        "title": "Gemini 3.5 Pro GA and first independent benchmark",
        "why": "Google's Pro has slipped to the edge of its June commitment. A GA with an independent Artificial Analysis pass will show whether it lands above or below Fable 5 and GPT-5.5."
      },
      {
        "window": "Jun-Jul",
        "title": "MiniMax-M3 weights release",
        "why": "If MiniMax ships the weights at its announced 55 AA Index, it becomes the leading open-weights model and a credible self-host alternative to closed frontiers."
      },
      {
        "window": "Jun-Aug",
        "title": "Independent SWE-Bench Pro replication of Fable 5",
        "why": "Fable 5's 80.3% is a vendor-harness figure. An Epoch or Scale SEAL replication will confirm or deflate the coding-lead claim that buyers are being asked to act on."
      }
    ],
    "changelog": [
      "Added Claude Fable 5, DiffusionGemma, and North Mini Code to the LLM tree; reframed W24 around a frontier launch-and-takedown and an open text-diffusion efficiency path."
    ]
  }
}