← Models & compute

Models & compute

July 2026

• Jul 31, 2026 Model releases In the briefing →

  OpenAI — GPT-5.6 Luna now costs $0.20 / $1.20 per million input and output tokens, a fifth of its launch price, with Terra down 20%. What paid for it is the odd part: OpenAI says GPT-5.6 Sol rewrote its own production serving kernels in Triton and Gluon, taking roughly 20% out of end-to-end serving cost. Luna lands at a fifth of Claude Haiku 4.5's input price. (official, source)

• Jul 31, 2026 Model releases In the briefing →

  Google DeepMind — Gemini Robotics ER 2 ships to developers through the Gemini API and AI Studio rather than a robotics-only preview. It plans multi-step physical tasks, watches video to confirm a step actually finished before moving on, and coordinates more than one robot. Google reports gains over ER 1.6 on safety-instruction-following and human-proximity benchmarks. (official)

• Jul 30, 2026 Model releases In the briefing →

  Anthropic — Claude went down globally, with requests failing on 529 Overloaded across the API, Claude Code and the apps. Anthropic opened an investigation at 19:49 UTC and said it had identified the cause at 20:33 UTC without naming it publicly. Service came back the same evening, still with elevated latency. (source)

• Jul 29, 2026 Model releases In the briefing →

  Liquid AI — Two open-weight encoders landed for the jobs nobody needs a frontier model for. LFM2.5-Encoder-230M and -350M handle 8,192-token context and target classification, routing, extraction and scoring; the 350M ranks fourth of fourteen models tested across GLUE, SuperGLUE and multilingual tasks, and the 230M beats ModernBERT-base and every EuroBERT variant. On CPU at full context the 230M takes about 28 seconds per forward pass against ModernBERT-base's minute and a half. (official)

• Jul 28, 2026 Model releases In the briefing →

  Moonshot — Kimi K3's weights are out, and the license is the part to read first. The model card lists 2.8 trillion parameters, 104 billion active (16 of 896 experts per token) and a 1M-token context, with 93.5% on GPQA Diamond and 88.3% on Terminal-Bench 2.1. The download runs 1.56TB, so self-hosting is a datacenter question. Moonshot calls this open-weight rather than open source: model-as-a-service vendors above $20M annual revenue must negotiate separate terms. OpenRouter resells it at $3/$15 per million. (official, source, source)

• Jul 28, 2026 Model releases In the briefing →

  Microsoft — Its first in-house security model arrived with a caveat from its own CEO. MAI-Cyber-1-Flash hunts vulnerabilities in large codebases, and Mustafa Suleyman says it beats Gemini, GPT-5.5 Cyber, GPT-5.6 Sol and Mythos 5 on Cyber Gym. The hardest cases still route to GPT-5.4 inside Microsoft's MDASH harness. A companion agentic system, Perception, runs red, blue and green agent teams and enters preview November 3. (source)

• Jul 28, 2026 Model releases In the briefing →

  Anthropic — Dario Amodei answered the open-weights letter: Anthropic has never advocated a ban, and he calls open weights a public good where the capabilities aren't dangerous. His asks are chip export controls, a formal crackdown on distillation, and a global model-safety-testing body that China would join. (source, source)

• Jul 27, 2026 Model releases In the briefing →

  Moonshot — Kimi K3's weights land at 15:00 UTC today, not overnight. The official repo at moonshotai/Kimi-K3 still renders an upcoming-release countdown, and its page data carries a releaseDate of 2026-07-27T15:00:00.000Z — so anyone who queued a download against a midnight drop, or read one of the "available now" headlines that ran overnight, is pulling nothing yet. What Moonshot has published so far is the July 16 announcement: 2.8T parameters, a 1M-token context, native vision. The licence text, the model card, and the technical report all arrive with the checkpoint, which means the modified-MIT terms and their reported monthly-active-user clause stay unverified for another few hours. (repo, vendor, background)

• Jul 26, 2026 Model releases In the briefing →

  Industry — The open-weights letter picked up the two labs that were reported missing from it. "Open Weights and American AI Leadership" asks Washington to expand compute access, fund shared training assets, and avoid premature restrictions on open models. Reporting on July 24 counted 25 signatories and named OpenAI, Google, and Anthropic as absent; the letter as published by NVIDIA and Microsoft lists 50 organizations, Google and OpenAI among them, alongside GitHub, Meta, Mistral, Hugging Face, Ollama, Cloudflare, and Palantir. Anthropic is the one frontier lab still off it. The trigger is a distillation fight — TechCrunch reports the White House has accused Moonshot of distilling Anthropic's Fable to build Kimi K3. (letter, Microsoft, source)

• Jul 26, 2026 Model releases In the briefing →

  OpenAI — ChatGPT, Codex, and a dozen API endpoints went down worldwide for about 50 minutes starting 5am ET, with requests failing on "too many concurrent requests" errors. OpenAI acknowledged it on the status page at 5:30 and had everything back by 6. Brief, but the coding agent and the API went together. (source)

• Jul 26, 2026 Model releases Weekly recap Read the recap →

  Opus 5 at half Fable 5's price, Google's 3.6 Flash trio without a 3.5 Pro, and an open-weights letter that turned into a distillation fight.

• Jul 25, 2026 Model releases In the briefing →

  Anthropic — Claude Opus 5 ships as the new default Opus: roughly Fable 5-level intelligence at half Fable's price, with a per-request effort dial. Anthropic released Claude Opus 5, live across Claude.ai, the API, Claude Code, and Cowork, now the default on Claude Max and the strongest model on Pro. Standard pricing is $5 / $25 per million input/output tokens — identical to Opus 4.8 and half of Fable 5's $10/$50 — with a fast mode running ~2.5× faster at double the base rate ($10/$50); the context window is 1M tokens. The headline feature is a per-request effort control (low/medium/high, extending to xhigh/max) that trades token spend for reasoning depth, and Anthropic says Opus 5 converts extra effort into better results more reliably than any earlier Opus. On its own numbers it lands around Fable 5-level intelligence at half the price, sets a new state-of-the-art on agentic coding and knowledge-work benchmarks (Anthropic cites 96.0% SWE-bench Verified, 79.2% SWE-bench Pro, and a jump to 59.4% on SWE-bench Multimodal), and more than doubles Opus 4.8 on agentic terminal coding, though it trails the Mythos line on cybersecurity. Anthropic also frames it as its safest Opus: ~85% lower safety-classifier trigger sensitivity than Fable 5, the lowest misalignment score it has measured, and no data-retention requirement for general access (unlike Mythos-class models). It matters because it inverts the usual flagship pattern — a stronger frontier model arriving at the *same* Opus price and *below* Fable's, with a cost/capability dial — resetting what the top tier costs to run rather than raising it. (official, TechCrunch, Fortune, Bloomberg)

• Jul 23, 2026 Model releases In the briefing →

  Google — Gemini 3.6 Flash lands with two siblings — 3.5 Flash-Lite and a gated 3.5 Flash Cyber — but still no 3.5 Pro. 3.6 Flash is the new "workhorse" tier: Google says it uses ~17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index — up to 65% fewer on DeepSWE (by Datacurve) — at a lower cost per output token, while scoring higher on every internal eval. Artificial Analysis clocks the high-reasoning variant at 50 on its Intelligence Index (tier median 31) and ~275 tokens/sec, with pricing around $1.50 / $7.50 per million input/output tokens ($0.15/M cached) and the knowledge cutoff jumping from Jan 2025 to March 2026. It's live for developers today in the Gemini API via Google AI Studio and Android Studio. 3.5 Flash-Lite is the fastest, cheapest 3.5-class model (~350 tokens/sec) and is rolling into Google Search; 3.5 Flash Cyber, tuned to find and fix vulnerabilities, is deliberately gated to governments and trusted partners via CodeMender under a limited-access pilot, citing the tech's dual-use risk. It matters because after three missed 3.5 Pro targets and a "months behind schedule" report, this is the first concrete evidence Google's Gemini pipeline is shipping again — the "upgraded Flash could land first" scenario the last several briefings tracked — and a cheaper-yet-stronger Flash resets the cost/quality floor for the workhorse tier. (official, benchmarks, source)

• Jul 21, 2026 Model releases In the briefing →

  Anthropic — Fable 5 subscription terms finalized: permanently included on Max and Team Premium (at 50% of limits), metered for Pro and Team Standard. After a run of week-long extensions that kept Fable 5 temporarily included while Anthropic brought inference capacity online, the company settled the terms in an announcement ahead of the July 20 cutover: Max and Team Premium plans now include Fable 5 permanently, capped at 50% of each plan's weekly usage limits, while Pro and Team Standard lose included access — those users get a one-time $100 usage credit, after which Fable 5 bills at API rates of $10 / $50 per million input/output tokens. This resolves the watch-list item the last several briefings carried as an open question: the previously rumored, unconfirmed "permanent-50%-for-Max vs credits-for-Pro" split is now the confirmed, live billing structure rather than a placeholder deadline. Anthropic still frames the 50% cap as capacity-driven, not a permanent ceiling. It matters because it ends the six-week "will it be included?" limbo with a concrete two-tier answer — and it draws a sharp line between Max/Team Premium (Fable 5 is now a standing plan feature) and Pro/Team Standard (every session past the $100 credit is a metered line item). (The New Stack, official, prior coverage)

• Jul 19, 2026 Model releases Weekly recap Read the recap →

  Moonshot's Kimi K3 (2.8T/~50B-active MoE, 1M context, #1 Frontend Code Arena, weights promised July 27) was the week's one shipped flagship; Google's Gemini 3.5 Pro missed July 17 and was confirmed months-late on coding, taking ~$200B off Alphabet.

• Jul 18, 2026 Model releases In the briefing →

  Moonshot AI — Kimi K3 launches as the largest open-weight model ever announced and takes #1 on the external Frontend Code Arena, ahead of Fable 5 and GPT-5.6 Sol. Moonshot shipped Kimi K3 via its API (api.moonshot.ai, model id kimi-k3): a sparse Mixture-of-Experts model with 2.8 trillion total parameters but only ~50B active per token (896 experts, 16 routed), a 1M-token context window, and multimodal input (text, images, video). On Arena.ai's Frontend Code Arena — an external, human-preference leaderboard — K3 took #1 with 1,679 points, ahead of Fable 5 (1,631) and GPT-5.6 Sol (1,618), winning six of the board's seven domains and losing only Gaming to Fable 5. Moonshot's own benchmark table is more modest: it reports K3 topping Opus 4.8 and GPT-5.5 while still trailing Fable 5 and GPT-5.6 Sol on most suites, so the Frontend Code Arena result is a genuine external SOTA on that specific board, not an across-the-board lead. API pricing is $3 / $15 per million input/output tokens (cache-hit input $0.30) — roughly Sonnet-tier pricing for a model Moonshot frames as Opus-4.8-class. It matters because a Chinese lab now holds the top spot on a widely watched frontend-coding leaderboard with a model whose weights are slated to go public, landing on the exact coding axis where the US flagships compete — and it does so in the same window Google's Gemini 3.5 Pro slipped for coding shortfalls. The catch: the open weights aren't out yet — Moonshot says they release July 27, so today K3 is a proprietary API with an open-weights promise attached, and independent evals beyond the frontend arena are still thin. (source, analysis)

• Jul 17, 2026 Model releases In the briefing →

  Google — Gemini 3.5 Pro misses its July 17 target; Bloomberg reports the flagship is "months behind schedule," and Alphabet sheds ~$200B in market value. A day before the widely reported July 17 GA date, Bloomberg — citing people familiar with the matter — reported that Google is months behind on delivering Gemini 3.5 Pro, its most powerful model, because the company keeps taking more time to improve it, with coding capabilities in particular falling short of internal expectations. The reporting says the slippage has frustrated Google engineers, researchers, and managers who worry the company is ceding ground as Anthropic and OpenAI ship models that exceed Gemini's. Google did not announce a GA; asked about status, it said only that it is "currently testing 3.5 Pro, an upgraded Flash model, and other models with partners." Markets took the delay badly: Alphabet closed down ~4.4% on July 16, wiping out roughly $200 billion in market capitalization. It matters because the July 17 date has anchored this briefing's watch list for over a week as the last unlaunched major summer flagship — today it lands not as a launch but as a confirmed, open-ended slip, and the stated reason (coding, not just token efficiency) is the exact axis on which Fable 5, GPT-5.6, and Grok 4.5 are already competing. (source, report, prior coverage)

• Jul 13, 2026 Model releases In the briefing →

  Anthropic — Fable 5 stays included in Claude subscriptions through July 19; the metered cutover slips to July 20. Fable 5 was due to leave Pro, Max, Team, and select Enterprise plans at 11:59pm PT July 12, moving to usage credits; instead Anthropic extended the included window a second time — now 11:59:59pm PT July 19 — and carried the paired 50% boost to Claude Code weekly limits along with it. The extension applies automatically, with no action required from users. After July 19, Fable 5 moves to usage credits at $10 / $50 per million input/output tokens — the top of Anthropic's published rate card. Anthropic frames the metered move as temporary: a Claude Code lead engineer said the aim is to restore Fable 5 as a standard part of subscriptions once Anthropic has enough capacity, casting the switch as a demand/capacity constraint rather than a permanent pricing decision. It matters because the last two briefings led with this cutover as a hard July 12 deadline teams were triaging production paths against — the reprieve buys another week, but the load-bearing detail is that capacity, not price, is what now gates whether Fable 5 ever settles back into the included tier. (source, context, official, prior coverage)

• Jul 12, 2026 Model releases Weekly recap Read the recap →

  Three near-frontier models in one week: GPT-5.6's broad GA (Sol $5/$30, Terra $2.50/$15, Luna $1/$6), Grok 4.5 ($2/$6), and Meta's Muse Spark 1.1 ($1.25/$4.25). OpenAI also advanced voice — gpt-realtime-2.1 (reasoning on the Realtime API, −25% p95 latency) and the full-duplex GPT-Live, which replaces ChatGPT's Advanced Voice Mode.

• Jul 11, 2026 Model releases In the briefing →

  Meta — Muse Spark 1.1 launches on the paid Meta Model API — Meta's first time charging for its own model. Meta Superintelligence Labs shipped Muse Spark 1.1, a multimodal reasoning model built for agentic and coding work, and — for the first time — is charging developers to call Meta's own model through the new Meta Model API (US developer public preview, $20 in free credits). It carries a self-managed 1M-token context window and native primary-agent/subagent orchestration (it can plan and delegate across parallel subagents, or run as a subagent that escalates back), with MCP and custom-skill support and reported zero-shot generalization to new tools. On tool-use benchmarks Meta claims the lead — MCP Atlas 88.1 (scaled tool use) and JobBench 54.7 (professional tool use) — but on coding it trails the flagships: SWE-Bench Pro 61.5 (Opus 4.8 69.2), Terminal-Bench 2.1 80.0 (GPT-5.5 83.4, Opus 4.8 82.7), DeepSWE 1.1 53.3 (GPT-5.5 67.0, Opus 4.8 59.0). Pricing is $1.25 / $4.25 per million input/output tokens — roughly a quarter of Opus 4.8 ($5/$25) and GPT-5.5 (~$5/$30). It matters because it is a strategic reversal: Meta's model story was open-weight Llama you ran yourself, and Muse Spark 1.1 is a metered, agent-first API you rent — the third price-aggressive agentic model in a week after Grok 4.5 (July 8), but the first that puts a major open-weights lab behind a per-token meter. (official, source, enterprise)

• Jul 10, 2026 Model releases In the briefing →

  xAI/SpaceXAI — Grok 4.5 ships as a coding-and-agentic model trained in partnership with Cursor, priced far below the frontier. SpaceXAI released Grok 4.5 on July 8 (developers first, wider public rollout the next day), pitched at coding, agentic tasks, and knowledge work and trained on real Cursor developer-session data — not just code but the full prompt-iterate-refine workflow. Artificial Analysis places it 4th on its Intelligence Index at a score of 54, behind only Fable 5, GPT-5.5, and Opus 4.8 and a 16-point jump over Grok 4.3, with Terminal-Bench 2.1 at 83.3% (within a point of GPT-5.5 and Fable 5) but DeepSWE 1.1 at 53%, well behind GPT-5.5 (67%) and Fable 5 (70%). The headline is price: $2 / $6 per million input/output tokens (cached input $0.50), a 500K-token context window, and availability in Grok Build, Cursor on all plans, and the SpaceXAI console — EU access is not live yet, expected mid-July. It matters because it drops a near-frontier model at a rate Artificial Analysis measures as >60% below Opus 4.8 and GPT-5.5, extending the same cost pressure that pushed US buyers toward cheap open-weight models last week, now from a US lab with a coding-specialized model rather than an open-weight substitute. (official, analysis, source)

• Jul 9, 2026 Model releases In the briefing →

  OpenAI — GPT-5.6 (Sol, Terra, Luna) goes broadly available, ending the government-restricted rollout. OpenAI said on July 8 the models become publicly available starting July 9, after the US Department of Commerce cleared a wide launch under Washington's new frontier-model oversight framework — lifting the restriction that had kept GPT-5.6 to a ~20-partner, government-vetted preview since June 26. The tiers and preview pricing are unchanged: Sol (flagship, OpenAI's "strongest model yet," pitched as more capable across coding, biology, and cybersecurity) at $5/$30 per million input/output tokens, Terra (balanced, ~2× cheaper than GPT-5.5) at $2.50/$15, and Luna (fast, low-cost) at $1/$6. It matters because it closes the single biggest open item on this feed's watch list — a frontier model held back for weeks under a federal review process — and does so the same way Anthropic's Fable 5 return did in late June: a government clearance, not a capability milestone, was the gate. Watch for the first independent benchmarks now that Sol is out from behind the vetted-partner wall. (news, regulatory, preview)

• Jul 9, 2026 Model releases In the briefing →

  OpenAI — GPT-Live: full-duplex voice models replace Advanced Voice Mode. Alongside the GPT-5.6 news, OpenAI introduced GPT-Live-1 and GPT-Live-1 mini, a new voice generation built on a full-duplex architecture that listens and speaks at the same time — continuously processing input while generating output, so it decides many times a second whether to speak, keep listening, pause, interrupt, or call a tool. For questions needing search or deeper reasoning it delegates to OpenAI's frontier model behind the scenes and folds the result back into the conversation. It's rolling out to ChatGPT users globally, with GPT-Live-1 mini replacing Advanced Voice Mode by default and the larger GPT-Live-1 on paid tiers; an API is promised "soon." It matters because it's a different bet from the gpt-realtime-2.1 API models shipped July 6 (reasoning bolted onto a turn-based Realtime pipe): GPT-Live rearchitects the consumer voice stack itself around continuous listen-and-speak, moving ChatGPT voice off the transcribe→model→TTS pattern entirely. (official, source)

• Jul 8, 2026 Model releases In the briefing →

  OpenAI — gpt-realtime-2.1 and gpt-realtime-2.1-mini shipped on the Realtime API. The update brings reasoning to realtime voice — with configurable reasoning effort and a spoken preamble so the model can keep talking while it works through a multi-step request — and improves alphanumeric recognition, silence/noise handling, and interruption behavior; OpenAI says it cut p95 latency by at least 25% across its Realtime voice models through improved caching. The mini tier is a distilled reasoning model aimed at faster, lower-cost voice agents. It matters because it pushes the Realtime API past a transcription-and-TTS pipe toward voice agents that can reason and use tools mid-conversation, and pairs that with the latency and cost cuts that make always-listening agents viable in production. (official, docs)

• Jul 5, 2026 Model releases Weekly recap Read the recap →

  the export lift restored Fable 5 (July 1, behind a new cyber classifier and 50% cap) and Mythos 5; Sonnet 5 shipped with a native 1M context at $2/$10; Claude reached GA in Microsoft Foundry; and Claude Science opened as Anthropic's first industry-vertical product.

• Jul 3, 2026 Model releases In the briefing →

  Anthropic — Claude Science shipped, a dedicated AI workbench for scientists that unifies 60-plus scientific databases across genomics, proteomics, and cheminformatics into a single environment where researchers can work with dozens of file formats and analysis tools "at all stages of their work." Currently in beta, it lands alongside Anthropic's own move into in-house drug-discovery programs — with a stated early focus on neglected diseases that major pharma has passed over on profitability grounds — putting Anthropic into a three-way AI drug-discovery race with Google and OpenAI. Early customers include Novo Nordisk and the Allen Institute, and Anthropic is offering up to $30,000 in use credits plus project funding, with applications open through July 15. It matters because it is Anthropic's first vertical product built for a specific industry rather than a general model or coding tool — a bet that the near-term enterprise pull for frontier models runs through domain-specific research workflows, not just chat and code. (official)

• Jul 2, 2026 Model releases In the briefing →

  Anthropic — Fable 5 and Mythos 5 are back after the US lifted export controls. On June 30 the Commerce Department lifted the June 12 export-control directive that had forced Anthropic to suspend both frontier models globally, ending a 19-day blackout; Anthropic said it received notice and would begin restoring access the next day. Fable 5 returned July 1 across Claude.ai, the Claude Platform, Claude Code, and Claude Cowork, with AWS, Google Cloud, and Microsoft Foundry re-enablement to follow "as soon as possible." It comes back as a battened-down build behind a new cybersecurity safety classifier — trained in coordination with the government — that targets and blocks the safeguard-bypass behavior (a jailbreak Amazon researchers surfaced) that triggered the original order. Access terms are temporary: Pro, Max, Team, and select Enterprise plans get Fable 5 at up to 50% of weekly usage limits through July 7, after which it's available via usage credits. It matters because the last frontier piece is restored — and unlike Mythos 5's June 26 org-gated return, Fable 5 comes back as a general unlock, but with a new model-level cyber classifier now part of the cost of shipping it. (official, CNBC)

• Jul 2, 2026 Model releases In the briefing →

  Anthropic — Claude Sonnet 5 shipped, a new mid-tier model with a native 1M-token context window, positioned as "the most agentic Sonnet yet" and close to Opus 4.8 quality at a lower price. Introductory API pricing is $2 / $10 per million input/output tokens through August 31, then $3 / $15. It's the default model on Free and Pro (available to Max, Team, and Enterprise) and is already the default in Claude Code (v2.1.197). On Anthropic's launch benchmarks it beats Sonnet 4.6 across the board (reported SWE-bench Pro of ~63% vs ~58%), landing between Sonnet 4.6 and Opus 4.8 (~69%) on agentic coding. It matters because it widens the cost-performance menu — near-frontier agentic behavior at a mid-tier price, with a 1M context Sonnet didn't previously have. (official)

• Jul 1, 2026 Model releases In the briefing →

  Anthropic — the export controls on Fable 5 and Mythos 5 have been lifted, ending the blackout that took both models offline on June 12 (day 18). Commerce Secretary Howard Lutnick told Anthropic it no longer requires an export license after the company agreed to "proactively detect and address security risks associated with the models," to work with the government on protocols for future releases, and to report any "malicious activity" it finds. Anthropic says it will begin restoring access on July 1, with Fable 5 returning to users globally; it did not spell out whether consumer access comes back all at once or through a gradual rollout. This completes the reversal that began June 26, when Commerce first cleared the stronger cyber model, Mythos 5, for ~100 critical-infrastructure defenders — the full lift now covers the consumer-grade Fable 5 as well. Alongside the statement, Anthropic said it is proposing an industry-wide framework for scoring jailbreak severity together with Amazon, Microsoft, Google, and other Glasswing partners — a direct answer to the "narrow jailbreak" finding that triggered the June 12 suspension. It matters because the month-long saga of state-gated frontier access — the throughline of nearly every briefing this month — resolves with the models returning under a negotiated compliance regime rather than a fresh directive, and with a proposed standard meant to keep a single reported jailbreak from pulling a deployed model again. (official, CNBC, 9to5Mac)