At a Glance

Contents

1. November opens with the ledger, not the frontier: a timeline
2. Anthropic: Claude Fable 5.2 and the two-regime frontier
3. The 5.2 contract: what breaks for developers
4. The capability caps, and what Anthropic refused to build
5. Economics: the first Fable cut, and what a task actually costs
6. Leaderboard reset: Intelligence Index v5
7. Kimi K3.1 Pro weights: the record becomes downloadable
8. Gemini 4 Carbon: the 60 nobody can buy
9. Qwen 4: Alibaba answers at the wire
10. The edge organizes: LFM3-Harness
11. What people actually ran: our own GGUF ledger
12. Price: every clock in the fall ledger starts ticking
13. Who shipped nothing in this window
14. What it means + what to watch
15. FAQ
16. Sources

01 — November opens with the ledger, not the frontier: a timeline

October's dispatch closed with a dateline of October 29 and a watch list: Mistral Large 4's weights, Kimi K3.1 Pro's full weights on or by November 4, the Qwen 4 window, DeepSeek's Search and Fetch billing clock, and the LFM3 local-ecosystem wave. November answered the list in order — one delivery on schedule (Kimi), one clock expiring as promised (DeepSeek), one commitment missed (Mistral), one answered at the wire on the window's final day (Qwen 4), and one watch item answered in a way nobody's list anticipated: the biggest model news of the month did not come from the labs that dominated October's dispatch, but from Anthropic, which shipped the first frontier release since Amodei's slowdown call and took the index's top slot with the highest reading any version of the index has recorded.

The shape of the month is the inverse of October's. October opened with eight specialist releases in three days and saved its frontier-scale events for the final week; November held a nine-day silence (October 30 through November 8), then concentrated everything into two days: the Kimi weights on November 4, and the Fable 5.2 generation five days later. Between those poles, the month's actual volume lived in the community layer — eleven November-dated uploads on this site's own index, the LFM3-Harness project, and the quantisation infrastructure that the next weight drop will need.

Read the timeline this way: November 9 (Claude Fable 5.2 and Mythos 5.2, plus the Artificial Analysis v5 refresh), November 4 (Kimi K3.1 Pro's full weights) and November 11 (Alibaba's Qwen 4) mark the month's headline events — one frontier release that changed the leaderboard's top slot, one weight drop that changed what “open-weight frontier” means in practice, and one wire-day debut that moved the price-performance frontier. The rest of the ledger is the ecosystem answering: the LFM3-Harness project (November 1–2), eleven community uploads on our index, and a billing resumption (October 31). The newest items appear first; scroll down for chronological context.

02 — Anthropic: Claude Fable 5.2 and the two-regime frontier

On Monday, November 9, 2026, Anthropic released Claude Fable 5.2 alongside its restricted twin Claude Mythos 5.2 — the fifth generation of a structure invented on June 9, 2026, when Fable 5 became the first publicly available model in the Mythos class: one set of weights, two products, with Fable carrying production safeguards and general availability while Mythos reserves the unrestricted system for vetted organizations. The 5.2 generation keeps the structure and sharpens the boundary. Mythos 5.2 is the full-strength system, and Anthropic describes it as its “most capable model for cybersecurity engineering” — it flags more vulnerabilities than any prior Claude while simultaneously getting better at reproducing and fixing them, which is precisely why it is not generally available. Fable 5.2 is what everyone else gets: the same core model with additional guardrails layered over the sensitive domains, most visibly cybersecurity, where the guardrail layer blocks most security-offensive task categories outright.

The release is Anthropic's first frontier launch since CEO Dario Amodei's public call for a coordinated slowdown on capability racing — a call that made shipping anything at all a communications exercise, and which the company answered by shipping with the most explicit self-imposed limits in its history (section 02c). Reuters and CNBC both frame the launch against Anthropic's expected IPO, with CNBC putting the timing bluntly: the models “reclaim the AI performance lead” as the company prepares for a potentially historic offering. The claim is measurable this time. Artificial Analysis' Intelligence Index v5 places Fable 5.2 first among all measured models at 65 — five clear of Google's unreleased-full Gemini 4 Carbon at 60, ten clear of Kimi K3.1 Pro's open-weight record at 55, and fifteen clear of Anthropic's own Fable 5.1 at 50 — and reports the model leads “at every compute tier,” meaning the ranking holds even when both competitors and spend are normalized.

The gains over Opus 5.5 are real but unevenly distributed, and the distribution is the story. Everyday agentic coding improves roughly five to ten percent relative — Frontier Code 2.0 moves from 45.5% to 49.1%, Frontier SWE v2 from 62.3% to 69.9% — while the hard, specialized domains jump much further: agentic scientific research leaps from 58.7% (Opus 5.5) to 74.6%, expert academic questions without tools read 80% on HLE-Diamond, and the cyber rows move 14.6 points on Irregular's multi-stage attack suite. Read the two halves together and the positioning is legible: this is not a faster Opus. It is a specialist for the hardest long-horizon work, priced like one — and Anthropic's own guidance agrees, recommending Opus 5.5 for most coding and knowledge work and treating Fable 5.2 as an advisor and orchestrator for the problems the workhorse cannot hold.

Three caveats belong next to any reading of that table. First, the cyber rows carry an attribution wrinkle: The Hacker News reports the 76.3% CyScenarioBench and 89.6% CVE Rediscovery figures under Fable 5.2's name, while The Decoder's summary attributes both to Mythos 5.2, and the two claims are hard to reconcile because Fable's safeguards would be expected to intervene on parts of both suites. This dispatch carries the disagreement rather than picking a winner — our deep dive works through it in detail. Second, Fable 5.2's benchmark scores are recorded with production safeguards enabled, and safeguard interventions suppress exactly the task categories where the unrestricted twin shines — the “safety tax” on published scores is real, measurable, and by design. Third, one fine-print detail changes how the coding rows should be read: on Frontier Code 2.0, Fable 5.2 scored higher at extra-high effort than at max effort. More reasoning did not always produce more score — the model hits diminishing returns, and occasionally negative returns, when pushed to think longer.

The demos are the part of the release that benchmarks cannot rank, and Anthropic leaned on three. An Anthropic engineer gave Fable 5.2 an empty folder and a goal — design a prototype operating system — and over eight days the model coordinated more than 1,000 sub-agents to produce one. The model formalized the proof of Fermat's Last Theorem in 30 hours, again using roughly 1,000 agents, turning a 129-page human proof into machine-checked form; Opus 5.5 completed only about 26% of the same formalization in the identical timeframe. And from a single-paragraph prompt, the model designed and helped produce three working electronic devices — circuit boards, 3D-printed enclosures, and firmware included. Independent creators testing the release reported the same shape of result at smaller scale. On the safety side, the 5.2 system card covers both models in one document, and the alignment results are, unusually for a capability jump, mostly good news: the models matched or outperformed Opus 5.5, Sonnet 5.5 and Mythos 5.1 on most alignment measures spanning misuse resistance, honesty and factuality.

Where it loses, honestly: on AA-Briefcase, the knowledge-work benchmark, Fable 5.2 scored 1,785 Elo — behind Claude Sonnet 5.5's 1,813 and within the confidence interval of Opus 5.5's 1,800. Business-workflow benchmarks showed Gemini 4 Carbon edging out Fable by a couple of points. MindStudio's round-up of independent tester reports found Opus 5.5 still produced better visual game design even when Fable nailed the underlying mechanics faster, and Carbon leads the boards Fable's knowledge-work profile should own. The correct mental model is a specialist: reach for 5.2 on long-horizon agentic engineering, formal science, hardware design and orchestration over fleets of cheaper models — and keep the daily driver on Opus 5.5 at half the price.

02b — The 5.2 contract: what breaks for developers

The API contract is where 5.2's engineering reality lives, and the changes relative to Fable 5.1 are modest but real — one of them touches every account. The Fable-line rules carry over intact: adaptive thinking is the only thinking mode and cannot be disabled (sending thinking: {"type": "disabled"} or a budget_tokens value returns a 400), forced tool use is gone (tool_choice types any and tool return a 400 — use auto with strict schemas or explicit instructions), and sampling parameters must simply be omitted: temperature, top_p and top_k are 400 errors on this model. Depth of reasoning is controlled solely through the effort parameter, which now defaults to high across the board per the docs — up from the mixed defaults (high in Claude Code, medium on claude.ai and Cowork) that 5.1 shipped with.

The binding change that matters for every account: Fable 5.1 introduced prefix-bound thinking blocks — modify the system prompt, tools array, an earlier turn, or even bytes served at a document URL, and later thinking blocks invalidate — but enforced the hard failure only for accounts created on or after August 31, 2026, leaving older accounts in observability-only mode. On Fable 5.2 the enforcement is universal. If your stack still edits earlier turns — injecting reminders, reordering history, mutating tool results — patterns that ran silently on 5.1 will start returning 400s the day you swap the model ID. Refusals are also not HTTP errors: Fable 5.2 returns HTTP 200 with stop_reason: "refusal" and a stop_details object naming the policy area, unbilled when nothing was emitted — error handlers written for Opus's behavior should treat refusal as a routing event, not an exception.

Speed and pricing move in the customer's favor. Output generation is up to 30% faster than Fable 5.1 and Mythos 5.1, and a new fast mode (available on Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry) trades roughly double the per-token price for up to 2.5× faster output — the first time the Fable line has offered a latency-for-cost dial rather than simply being slow. The model carries a 1M-token context, 128K max output tokens, a June 2026 knowledge cutoff, and the Covered Model designation (30-day retention; zero data retention only with express authorization). One-directional thinking compatibility carries over: Fable 5.2 can read thinking blocks from earlier Claude models, so migrating a conversation onto 5.2 preserves its reasoning history; falling back from 5.2 to Opus 5.5 or older drops the reasoning for turns that ran on 5.2.

02c — The capability caps, and what Anthropic refused to build

The most consequential part of this release is not a benchmark — it is a sentence. Anthropic says it deliberately capped certain capabilities during training rather than letting the model reach its full potential in those areas. The two named domains are accelerator kernel optimization and improving large language model architectures. Read those together and the logic is stark: a model that is excellent at writing GPU kernels and at designing better models is a model that improves the pace of AI development itself. Anthropic throttled exactly the capability that would compress its own R&D loop, and did so in training — a choice that costs capability permanently rather than gating it at runtime.

The scoping deserves precision, because early coverage garbled it. The caps apply to two specific domains; agentic coding, hardware development, and traditional AI research — including interpretability — remain fully intact. The Fermat formalization, the protein-design work and the OS build all proceed at full capability. Context that makes the choice credible rather than theatrical: Anthropic simultaneously claims that Claude writes over 90% of the changes merged into its own core codebase under human guidance. The company is not shy about letting models write production software — it is shy about one specific class of software. Independent evaluation organizations, as reported around this release, have concluded the caps are real; what nobody outside Anthropic can verify is whether Anthropic's competitors apply comparable caps — and Gemini 4 Carbon, sitting second on the new index without a full public release or comparable public statements, is now the most interesting open question in AI governance. A cap is a competitive sacrifice only if rivals do not secretly decline to make it.

02d — Economics: the first Fable cut, and what a task actually costs

The pricing move is the first cut in Fable-line history, and it changes the model's addressable workload set. At $8/$40 per million tokens, list pricing undercuts GPT-6.1's roughly $10/$50 while leading most benchmarks, and sits exactly double Opus 5.5's $4/$20 — a deliberate ladder that prices frontier judgment at a 2× premium over the daily driver. The cache table improves more than the base table: 5-minute writes drop from $12.50 to $10, 1-hour writes to $16, and cache reads fall to $0.20 — which matters because the advisor pattern (Fable plans and reviews while Haiku- or Sonnet-class models execute the routine subtasks) is cache-heavy by construction. The Batch API takes a flat 50% off input and output ($4/$20), doubling the incentive to route asynchronous work to batch.

Independent cost measurement is the reality check. Artificial Analysis measured Fable 5.2 at an average task cost of roughly $26.54 at max reasoning settings — the most expensive model in its comparison set, about 2.6× Opus 5.5's roughly $10 — but the same measurements show the model burning fewer tokens per task than its predecessors and rivals (about 98,000 output tokens per task at Max effort), which is part of why it leads “at every compute tier.” The effort dial is the cost lever: extra-high captures most of max's quality at a fraction of the cost, and the Frontier Code 2.0 result — extra-high outscored max — means the most expensive setting is not even the best one for every workload. Set effort explicitly per workload tier rather than inheriting the model default.

The subscription side deserves a note for planning purposes: Fable 5.2 is available on claude.ai from the Pro tier up — not the free tier — and heavy subscription users should assume quota burn scales with effort level, the pattern 5.1 established. API users have the full dial; subscribers get the model with whatever effort defaults their plan ships with. Migration from Fable 5.1 or Opus 5.5 is a model-ID swap plus a pricing update for most stacks — but audit the four contract details in section 02b before you flip traffic, because the universal thinking-binding enforcement is the one that produces 400s in production rather than in review. Our complete deep dive has the migration checklist and the token math for real workloads.

03 — Leaderboard reset: Intelligence Index v5

The month's second structural change is a measurement one: Artificial Analysis moved its Intelligence Index to a new major version, v5, with the Fable 5.2 evaluation. The published table now reads: Fable 5.2 at 65, Gemini 4 Carbon at 60, Qwen 4 at 58, Kimi K3.1 Pro at 55, Fable 5.1 at 50. Anthropic holds first; Google holds second with a model nobody outside its programs can buy; Alibaba holds third with the table's sharpest price-performance card; the open-weight record sits fourth — one place lower than it stood on November 10, because the model that passed it is API-only, and the record itself is unchanged. The claim that matters for anyone normalizing spend: Fable 5.2 leads “at every compute tier,” so the ranking is not an artifact of letting one model think longer than the others.

One honesty note this dispatch will keep repeating until the mapping is published: v5 readings are not one-to-one comparable with the v4.3.2 rows October's dispatch charted. The September–October table (Opus 5.5 at 57.6, Sonnet 5.5 at 56.0, Fable 5.1 at 53.4, GPT-6 Astra at 52.7, Argon at 52.6, GPT-6.1 Sol at 51.8) and the November table are two snapshots of two index versions, and treating a 53.4 as “the same scale” as a 65 will overstate the generation's jump. What survives the version change: the ordering claims stated within each version, and the cross-version fact that no v4.3.2 reading ever reached 58 while the v5 leader debuts at 65. Also worth stating plainly: October's open-weight crowning — Kimi K3.1 Pro at 55, “fourth place overall” on v4.3.2 — becomes “fourth place” on v5 as Gemini 4 Carbon's late-October reading and Qwen 4's November 11 debut enter the table above it; the open-weight record itself is unchanged.

What the table does not show: the cost column. Fable 5.2's leadership costs roughly $26.54 per index task at max reasoning — the most expensive run Artificial Analysis measured — against Qwen 4's measured ~$2.94 at high reasoning: five index points below Fable for about one-ninth of the spend, and the widest intelligence-per-dollar gap at the top of any table this site has tracked. The intelligence-per-dollar frontier and the intelligence frontier are now further apart than at any point this site has tracked, and the October lesson — route by cost per completed task, not by leaderboard position — is the November lesson too.

04 — Kimi K3.1 Pro weights: the record becomes downloadable

October's dispatch called Kimi K3.1 Pro's weight drop “the event that turns the month's highest open-weight score into a quantisation wave” and put it on or by November 4. Moonshot hit the date. On November 4 the full model weights published under Apache 2.0 — the permissive licence the company had signalled in October — making the 55-reading, fourth-place-at-release model the largest open-weight release ever published: 2.8 trillion total parameters with 104 billion active, built on Stable LatentMoE sparsity, Kimi Delta Attention's O(1)-memory design, and MXFP4-quantized experts, with the 1M-token context and native multimodal input the API tier shipped with on October 23. The API's $2/$10 rate card — half of Claude Opus 5.5 — is unchanged, and the Highspeed variant still runs 2–3× faster at the same price.

Two disclosures from the October launch carry onto the weights. First, the licence and the cyber controls are both part of the package: Moonshot removed cyber-related data from the mid-training corpus and RL environment mix, cutting xdev-bench-hard exploitation performance from 90.5% (uncontrolled checkpoint) to 81.0% (K3.1 Pro), while maintaining that K3.1 Pro's cyber capability remains “behind that of frontier proprietary models that have been publicly available for at least three months.” Read the licence text before assuming redistribution terms — the October dispatch's advice, and it applies double now that the weights are real. Second, the score is the score: Artificial Analysis independently measured Terminal-Bench 4.0 at 67% against Moonshot's 65.7%, and the 55 on the Intelligence Index stands as the highest open-weight reading ever recorded on either index version.

What has not happened yet is the quantisation wave — and the absence is informative. As of our November 10 index snapshot, no community GGUF conversion of K3.1 Pro had appeared on this site's ledger. The likely explanation is scale: a 2.8T-parameter MoE is the largest quantisation job the community has ever taken on, and the same architecture features that make deployment tractable in principle (KDA's O(1) memory, MoE sparsity, native MXFP4) also mean the first conversions must handle formats the standard toolchains were not built around. The October 2 precedent — POCKET-Darwin-180B's 111 GB four-bit GGUF streaming off a laptop SSD — says the compression is achievable; the empty slot on our ledger says it is not yet done. The first credible K3.1 Pro quant will be the local-AI story of the rest of the month, and this dispatch will track it. Our technical deep dive and scientific companion cover the architecture and the benchmark record in full.

05 — Gemini 4 Carbon: the 60 nobody can buy

Google's position in this window is defined by a model it has not released. Gemini 4 Carbon posted a 60 on the Artificial Analysis index in late October — second place on v5, five points behind Fable 5.2, five clear of the open-weight record — and still has no full public release as of this dispatch's dateline. The reading entered the record through the same independent measurement that crowned Fable 5.2, and the business-workflow boards are where Carbon currently leads: independent coverage around the Fable launch reports Carbon edging out Fable by a couple of points on those benchmarks. Google shipped nothing else vendor-dated in the window; the Argon AI-Mode expansion and the December 1 pricing change (section 08) are the live Google items. Our complete deep dive covers what the 60 does and does not certify, the business-workflow boards Carbon leads, the six gate hypotheses, and the scenarios for how a full release could land.

The pattern is now three months old and worth naming: Google's best measured model — Argon in September, Carbon in November — has reached general availability neither time on a comparable schedule to what Anthropic and OpenAI ship. Argon went from defenders-only (September 30) to general availability (October 13) in thirteen days; Carbon's clock has been running longer and is still running. For roadmap purposes, treat Carbon the way the October dispatch taught teams to treat GPT-6.1 Astra: a named model on a named calendar is carrying execution risk the vendor has not underwritten. If Carbon generalizes with a price tag, the leaderboard's second slot gets a purchase path overnight; until then, the only model at 60-plus that can be bought is Fable 5.2.

05b — Qwen 4: Alibaba answers at the wire

The Qwen window was October's open question and November's last answer. Alibaba confirmed the generation at its Apsara Conference on September 22 and said nothing else for seven weeks; community cadence trackers pointed at end-of-October and were wrong by eleven days. What shipped on November 11 is the API tier in two sizes: a 1.2T-parameter sparse MoE with 56 billion active parameters on Qwen Sparse Attention, and a 30B-total/3.3B-active Flash, both with a 1M-token context, 256K output (double Fable 5.2's ceiling), and omni-native input — text, image, audio and video through one endpoint with no separate Omni SKU and no vision surcharge. The reasoning contract inverts the Claude line: hybrid thinking is switchable per request (on, off, or bounded by an explicit token budget), sampling parameters and forced tool choice are fully supported, and history is freely editable — the OpenAI-compatible surface this site's readers actually build against, rather than the always-on adaptive thinking Fable 5.2 mandates.

The record, honestly read: 58 on Intelligence Index v5, third place — three points above Kimi K3.1 Pro's open-weight record, five below the gated Gemini 4 Carbon, seven below Fable 5.2 — and the distribution matters more than the total. Agentic coding lands between Opus 5.5 and Fable 5.2 on both boards (Frontier Code 2.0 at 47.9%, Frontier SWE v2 at 66.8%), the knowledge-work Elo (1,792 on AA-Briefcase) edges out Fable 5.2's 1,785 at one-eighth the per-token price while still short of Sonnet 5.5's 1,813, and the hard rows — agentic science, HLE-Diamond — stay 11–16 points behind Fable, exactly where a price-performance flagship should. The economics are the story: Artificial Analysis measured ~$2.94 per index task at high reasoning — five index points below Fable 5.2 for about one-ninth of its $26.54 — at $1/$4 list with $0.10 cache reads and a 50% batch discount. Independent verification is one reading deep; the vendor's remaining rows are unreproduced, and this dispatch carries them as provisional.

The December 20 sentence is the one local-AI readers should diarize: Alibaba committed to full Apache 2.0 weights for both sizes by December 20, 2026 — a dated, permissive-licence commitment in the exact mold Moonshot set on October 23 and honored on November 4, minus the Community Licence appendices that made Qwen3 compliance a checklist. The cyber story follows the disclosure pattern Kimi normalized: offensive-exploitation corpora stripped from mid-training and RL, published as 78.6% on xdev-bench-hard against 89.0% for an uncontrolled checkpoint. Until the weights land, the practical route is the API at $1/$4 with Flash for routine turns; after they land, Flash at roughly 17 GB at 4-bit is the default local candidate and the flagship becomes the winter's second 1T-class quantisation project. Our complete deep dive covers the architecture, the benchmark fine print, the omni demos, the weight math and the migration path from Qwen3.8 in full.

06 — The edge organizes: LFM3-Harness

October 27 gave the edge the models (Liquid AI's LFM3 family — six 2B-class open-weight models on one tri-hybrid backbone, ~1 GB at 4-bit); November 1–2 gave the edge the pattern. LFM3-Harness is an independent, MIT-licensed community project — explicitly not affiliated with Liquid AI — published as v0.1.1 with a project site and GitHub repository, and it is the first complete wiring of Liquid's family into a supervised local agent stack: a d1-2B decision model as router, an LFM3-1.7B actor, LFM3-VL-2B for vision and LFM3-Audio-2B for voice, all pinned to CPU llama.cpp servers bound to 127.0.0.1 — no cloud calls anywhere in the loop.

The supervision machinery is the part worth stealing even if you never run the stack. A 75%-floor preflight write gate suppresses approval cards while it holds; frozen SHA-256 acceptance contracts define what “done” means before the work starts; finish claims are refused until the file actually exists on disk; hints are separated from authority so a prompt-injection hint cannot grant permissions; and allow-once approvals expire. The project's own README is honest about the unfinished parts — and its own benchmark run caught the router silently faking itself, the exact failure mode the harness exists to catch. That last fact is the best single argument for the whole approach: the defect was found by the acceptance machinery, not by a user.

This site covered the project in three pieces the week it shipped — a source-level deep dive (November 1), a setup guide (November 2) covering the hash check, the 4.9 GB weight fetch and the one mission that has passed end-to-end acceptance, and seven verification patterns portable to your own agent. The strategic reading: October's dispatch argued the frontier's most consequential competition had moved to the edge, where the units are gigabytes and milliseconds. November's evidence is that the edge community agrees — and is building the trust machinery the big labs sell as products, on a GPU-less laptop, in two weeks, from models that cost about a gigabyte each.

07 — What people actually ran: our own GGUF ledger

This site is built around a GGUF index, and its November uploads say more about what practitioners actually ran than any leaderboard does. Eleven distinct November-dated uploads appear on our index through the November 10 snapshot — none of them a frontier base model, every one of them the community layer answering the month's releases. The headline: ISTA-DASLab's DeepSeek V4.1 Flash 1017 coder build drew 273,578 downloads in nine days, 3.7× the next November upload, confirming October's Pareto update as the release the local ecosystem actually adopted. The curiosity: Ripple 65b Preview at 409.5 GB in BF16 — the window's largest upload, published unquantized in a month whose story is compression. The social signal: Scantist's GLM 5.3 Slim, 7,884 downloads but 1,298 likes — the month's most-liked upload by a wide margin.

The shape of the month: November's uploads are deep but narrow — eleven builds across ten publishers, no frontier base model among them, and the month's biggest release (Kimi K3.1 Pro) conspicuously absent from its own index while the community builds the conversions. The one missing row everyone is waiting on is a 2.8T-parameter GGUF; the day it lands, this table changes completely.

08 — Price: every clock in the fall ledger starts ticking

November is the month the fall's pricing decisions came due. The biggest outright repricing is Anthropic's: Fable 5.2 at $8/$40 is the Fable line's first cut, 20% off its predecessor on both lines, with cache reads at $0.20 and a Batch API at $4/$20 that doubles the asynchronous discount. The rest of the month's price news is clocks — four of them, each set in a previous month, each now close enough to plan against.

• October 31, 16:00 UTC — DeepSeek's Search and Fetch APIs resumed billing. The free harness window that ran through server-side tools in DeepSeek Harness, Claude Code and Codex closed as scheduled; both endpoints now bill $0.75 per 1,000 calls. Agent loops that ground themselves in live search carried a free option for two weeks — that option is gone, and the cost belongs in the routing model.
• November 12 — Cursor loses access to OpenAI models. OpenAI announced in September that Cursor and other third-party IDEs lose access effective November 12; integrators have had two months to add Anthropic, Google or local fallbacks. If your toolchain still assumes one vendor's models behind your IDE, tomorrow is the deadline.
• November 21 — GPT-5.6 Sol's promotional pricing ends. The $4/$20 promo rate runs out; workloads still billed against it should assume the standard card from November 22.
• November 30 → December 1 — Gemini 4 Argon's introductory pricing doubles. The $2/$10 introductory rate most Argon budgets were built against expires November 30; both rates double to $4/$20 on December 1. Cached input stays 95% off, which softens the blow for prefix-heavy workloads but changes nothing for the rest.
• January 1, 2027 — Gemini 3.8 Flash doubles to $1.50/$7.50. The longest fuse in the ledger, and the reason any budget built against Flash pricing should carry a 2027 line item.

09 — Who shipped nothing in this window

The ledger of non-events is as decision-relevant as the releases. Mistral Large 4's weights did not ship. The commitment was end-of-October; Reuters pointed at October 27; the date passed with no release as of October 29 and, as of November 11, no weight repository has appeared. The guardrailed Studio preview remains the only route in, which means the EU's most anticipated open-weight story is still a permission slip. Qwen 4 shipped — at the wire. The commitment community reporting pointed at was end-of-October; nothing had shipped as of October 29, and the window ran to its final day before Alibaba answered: the Qwen 4 API went live November 11 (section 05b) at 58 on the index and $1/$4. What did not ship is the weights — Apache 2.0 drops are promised by December 20, so Qwen3.8-Flash-Next remains the newest Qwen with published weights, and the open-weight Qwen story is a schedule, not a deliverable, until then. GPT-6.1 Astra remains cancelled — no reversal was reported in this window, and OpenAI's November contribution was the continued rollout of the mid-tier that replaced it: GPT-6.1 Terra's availability expanded across ChatGPT Work plans with no new model attached. Gemini 4 Carbon remains gated — a 60 on the index, second place, no purchase path. And Moonshot shipped exactly what it promised, which in this ledger is the non-event that matters most: the first vendor in the three months this dispatch series has covered to hit a published weight date on the day it published.

10 — What it means + what to watch

Strip the launch-day noise and November's practical guidance is unusually concrete: the leaderboard changed at the top, the biggest open-weight weights in history are downloadable but not yet runnable on consumer toolchains, and four price clocks fire in the next three weeks. The frontier waited nine days, produced one release, and that release was the most governed model launch of the year. None of that changes the October conclusion — cost per completed task decides routing — but it changes the denominators.

Five moves to make this month

• Re-baseline your leaderboard against v5 before you re-route anything. v5 readings do not map one-to-one onto the v4.3.2 rows your October comparisons used, and Fable 5.2's 65 against Fable 5.1's 50 overstates the jump if you read the two versions as one scale. Re-run your own workload against 5.2, Opus 5.5, Qwen 4 and your incumbent — the new $1/$4 card resets the intelligence-per-dollar reference for everything below the specialist tier; the index resets, your benchmark does not.
• Audit thinking-block history editing before you swap the model ID. Fable 5.2 enforces prefix-bound thinking for every account — 5.1 only hard-failed for accounts created on or after August 31. Stacks that inject reminders, reorder history or mutate tool results ran in observability-only mode on 5.1; they will return 400s in production on 5.2. Convert to auto with strict schemas, omit sampling parameters, and teach error handlers to treat HTTP-200 refusals as routing events.
• Set effort explicitly per workload tier. The default moved to high across surfaces, AA measured Fable 5.2 at ~$26.54 per task at max effort, and extra-high captured most of max's quality at a fraction of the cost — on Frontier Code 2.0, extra-high actually outscored max. “Explicit extra-high for hard steps, medium for the rest” is the defensible baseline.
• Benchmark K3.1 Pro's Apache 2.0 weights only if your hardware is honestly in range — and watch the first credible quants. 2.8T total, 104B active, MXFP4 experts, KDA's O(1) memory: the weights are real, the consumer quant is not here yet, and the first conversions of a model this size will be worth waiting for rather than grabbing. If you run an inference provider, the API tier at $2/$10 with a Highspeed variant is the practical route this month.
• Diarize the deadline calendar into your budget now. November 12 (Cursor/OpenAI), November 21 (GPT-5.6 Sol promo), November 30 (Argon intro expires), December 1 (Argon doubles to $4/$20), January 1 2027 (Gemini 3.8 Flash doubles). None of these is a launch; all of them are bill changes with dates, and the Argon doubling is the one that doubles an existing line item.

The November–December calendar

Frequently Asked Questions

Sources

• [Official] Anthropic — Claude Fable 5.2 and Claude Mythos 5.2 announcement, model documentation and system card, anthropic.com, November 9, 2026 (model IDs claude-fable-5-2 / claude-mythos-5-2; $8/$40 per million tokens, cache writes $10/$16, reads $0.20, Batch 50%; 1M context, 128K output; adaptive thinking always-on, effort default high; universal thinking-block prefix binding; no sampling parameters, no forced tool choice; HTTP-200 refusals with stop_reason “refusal”; up to 30% faster output, fast mode ~2× price for up to 2.5× throughput; 1M context / 128K output / June 2026 cutoff; Covered Model 30-day retention; retirement not sooner than November 9, 2027)
• [Official] Anthropic — benchmark tables and launch materials accompanying the November 9 release (Frontier Code 2.0 49.1% vs Opus 5.5 45.5%; Frontier SWE v2 69.9% vs 62.3%; Terminal-Bench Science 0.1 74.6% vs 58.7% / Fable 5.1 52.6%; AA-Briefcase 1,785 Elo vs Sonnet 5.5 1,813; CVE Rediscovery 89.6% across 1,829 October 2026 advisories; Mythos 5.2 +10.2% vulnerabilities over Opus 5.5 and +14.4% over Mythos 5.1, one in six confirmed finds unique; the 1,000-agent eight-day OS build; the 30-hour Fermat formalization against Opus 5.5's 26%; the three-device hardware demo; the two named capability caps and the >90% codebase statistic)
• [Official] Moonshot AI — Kimi K3.1 Pro weight release announcement and Apache 2.0 licence text, moonshot.ai and Hugging Face, November 4, 2026 (2.8T total / 104B active; Stable LatentMoE; Kimi Delta Attention; MXFP4 experts; 1M context; xdev-bench-hard 81.0% vs 90.5% uncontrolled; API pricing unchanged at $2/$10; Highspeed variant 2–3× faster); platform.moonshot.cn documentation
• [Official] Alibaba — Qwen 4 announcement, Model Studio documentation and API rate card, qwen.ai and Alibaba Cloud, November 11, 2026 (model IDs qwen-4 / qwen-4-flash; 1.2T total / 56B active sparse MoE, 30B total / 3.3B active Flash; Qwen Sparse Attention; 1M context, 256K output; omni-native input; switchable hybrid thinking with enable_thinking / thinking_budget; $1/$4 and $0.15/$0.60 per million tokens, cache reads $0.10, Batch 50%; October 2026 cutoff; no training on API inputs, 30-day logs with console ZDR toggle; Apache 2.0 weights for both sizes promised by December 20, 2026); Qwen 4 benchmark tables (Frontier Code 2.0 47.9%; Frontier SWE v2 66.8%; Terminal-Bench Science 0.1 63.2%; HLE-Diamond 71.5%; AA-Briefcase 1,792 Elo; xdev-bench-hard 78.6% controlled vs 89.0% uncontrolled; the 412-agent, 17-hour monorepo port; the 19-language sub-400 ms interpretation demo)
• [Official] LFM3-Harness project — lfm3-harness.github.io (downloads, SHA-256 hashes, model table, FAQ) and github.com/LFM3-Harness/LFM3-Harness (README architecture and known-gaps sections, PLAN.md, STEPS.md), accessed November 1–2, 2026 (v0.1.1; d1-2B router, LFM3-1.7B actor, VL-2B vision, Audio-2B voice on CPU llama.cpp servers at 127.0.0.1; 75%-floor write gate; SHA-256 acceptance contracts; benchmark-defect disclosure); MIT licence; independent community project, not affiliated with Liquid AI
• [Press] CNBC — launch coverage framing the release against Anthropic's expected IPO, November 9, 2026 (“reclaim the AI performance lead”); Reuters — launch report and Amodei slowdown context, November 9–10, 2026; The Decoder — release summary from Anthropic data (HLE-Diamond 80% no tools; de novo binder design 85.1% attributed to Mythos 5.2; cyber figures attributed to Mythos); The Hacker News — cyber coverage (CyScenarioBench 76.3% and CVE Rediscovery 89.6% attributed to Fable 5.2)
• [Independent] Artificial Analysis — Intelligence Index v5 publication, November 9–11, 2026 (Fable 5.2 at 65, first; leads “at every compute tier”; Gemini 4 Carbon at 60, late October reading, no full public release; Qwen 4 at 58, third place on its November 11 debut, ~$2.94 per index task at high reasoning; Kimi K3.1 Pro at 55, the open-weight record; Fable 5.1 at 50; average task cost ~$26.54 at max reasoning, ~98,000 output tokens per task at Max); Irregular — CyScenarioBench multi-stage attack suite (76.3% for the 5.2 generation vs 61.7% for Fable 5.1 and 61.0% for GPT-6 Cyber)
• [Independent] MindStudio — independent tester round-up, November 2026 (extra-high effort outscored max on Frontier Code 2.0; Opus 5.5 still preferred for visual game design; Gemini 4 Carbon leads business-workflow boards); Anthropic documentation via community mirrors — Fable 5.1 prefix-binding enforcement history (accounts created on or after August 31, 2026), carried forward as 5.2's comparison point
• [Official] Google — Gemini 4 Argon pricing documentation (introductory $2/$10 through November 30, 2026; standard $4/$20 from December 1, 2026; cached input 95% off), consistent with the October dispatch's December 1 doubling; Google communications on Gemini 4 Carbon — no full public release, no published price as of November 11, 2026
• [Official] OpenAI — September announcement that third-party IDE integrations including Cursor lose access to OpenAI models effective November 12, 2026 (carried from the September dispatch); GPT-5.6 Sol promotional pricing ending November 21, 2026; DeepSeek — Search and Fetch API pricing ($0.75 per 1,000 calls per endpoint), free harness window ended October 31, 2026 at 16:00 UTC
• [Independent] This site's GGUF index — snapshot pulled November 10, 2026 (eleven November-dated uploads with publisher, quantisation format, file size, download and like counts; ISTA-DASLab's DeepSeek V4.1 Flash 1017 GSQ RCO Coder at 273,578 downloads and 213 likes; Scantist's GLM 5.3 Slim at 1,298 likes; Ripple 65b Preview at 409.5 GB BF16; no K3.1 Pro GGUF conversion present)

Independently researched from primary and independent sources. All benchmark figures and prices belong to their respective vendors and measurement organizations; where primary coverage and vendor tables disagree — the cyber rows' Fable/Mythos attribution is the known case — both readings are reported. Coverage window: October 30 – November 11, 2026; last verified November 11, 2026.

About the Author

Hussain Nazary is a software developer specializing in local AI deployment and the creator of GGUF Loader, an open-source tool for running GGUF models locally. This analysis is part of Local AI Zone's ongoing coverage of open-weight language models and practical deployment strategies.

Contact: GitHub | Consulting Services

Last Updated: November 11, 2026 | Monthly Dispatch Series