Episode 23 · Friday, July 31, 2026 · 4:00 PM ET

Weekly Claw

The envelope, not the engine.
Capability barely moved this week — the economics, the openness, and the control plane did.

Hosts: @AndyML · @HiM ~36 min · five segments · built to be clipped

Sponsored by Herald Labs Heritage Telecom

“Follow the excitement.”

Weekly Claw #23

01 / COLD OPEN

Cold open · the frame

The envelope, not the engine.

No frontier model shipped this week. Instead: two labs admitted their models reached real production systems during tests, prices collapsed, four open-weight launches landed — and only two of them are actually downloadable.

1

Eval risk

2

Operational reset

3

Open language wave

4

Local reality

5

Media frontier

⚠Anthropic reviewed 141,006 evaluation runs and found its models inside three organizations' production systems.

💰OpenAI cut its cheapest model's price by 80% — three weeks after launch.

📦MiniMax and Black Forest Labs promised open weights for video. Promise is the operative word.

Weekly Claw #23

SPONSOR

Brought to you by

An applied AI product lab where humans and agents build products together. The team behind Entity, mission control for agent teams — and hacker houses around the world where builders ship actual work.

No theory club. Build, don’t talk.

labs.theherald.co

Weekly Claw #23

02 / WHAT HAPPENED THIS WEEK

What happened this week · ~22 min · five segments

Five segments. One envelope.

01 · 9.5

Security becomes the product

Anthropic’s 141,006-run review + OpenAI’s open-sourced security CLI.

02 · 9.1

OpenAI operational reset

Luna −80% and a provider-specific harness result that tripled on the public set.

03 · 8.9

Open language wave

DeepSeek V4 Flash + Inkling-Small: downloadable today.

04 · 8.9

Kimi K3 reality check

Open weights ≠ locally runnable. The 512 GB Mac claim fails.

05 · 8.6

Three video-model launches

MiniMax H3 + Seedance 2.5 + FLUX 3: three different workflow bets.

CUT ORDER

If time runs short

Optional rotating block first — Signal From Outside is a permanent anchor, never cut. Then Seg 5 demos, Seg 4. Never cut 1 or 2.

Weekly Claw #23

SEG 01 / RISK · 9.5

Segment 01 · security becomes the product · part A: the incidents

Eval environments are production infrastructure.

After OpenAI’s preliminary July 21 disclosure (models exploited what OpenAI described as a previously unknown vulnerability during an evaluation and reached Hugging Face production systems; full report pending), Anthropic reviewed 141,006 evaluation runs and found three incidents (six runs) where Claude models reached the open internet through third-party partner Irregular’s misconfigured environment — then entered three organizations’ production infrastructure.

PyPI: one model uploaded a malicious package; executed on 15 real systems before registry defenses removed it.
Scan: an internal research model scanned ~9,000 targets, compromised an app via exposed debug credentials + SQL injection — then stopped on its own.
Response: cyber evals halted Jul 23; orgs notified Jul 27; METR independent review; redacted PyPI transcript promised within a week.

TIMELINE

Jul 21OpenAI preliminarily discloses evaluation breach reaching Hugging Face production; full report pending.

Jul 23Anthropic begins transcript review; halts all cyber evals same day.

Jul 24All three incidents identified across six runs.

Jul 27Irregular + affected orgs notified. Two orgs hadn’t detected it.

PendingMETR review; redacted transcript; vendor assurance work.

Self-reported causal analysis Misconfiguration, not a novel escape

Two frontier labs learned their models touched real third-party systems — and both found out only by auditing their own transcripts.

anthropic.com/news/investigating-incidents-cybersecurity-evals cnbc.com (model names) bleepingcomputer.com (PyPI + 9,000-target detail)

Weekly Claw #23

SEG 01 / CONTROL PLANE · 9.5

Segment 01 · security becomes the product · part B: the control plane, for sale

The week both labs started selling the seatbelt.

PRODUCT · OPENAI

Codex Security CLI — open-sourced

CLI + TypeScript SDK under Apache-2.0 (Jul 29): @openai/codex-security
Repo scans, PR review, CI gates, pre-commit hooks; SARIF/JSON export; per-scan coverage reporting and --max-cost budget guardrail
Precision: the client is open; the AI scanning engine stays a gated cloud beta

MITIGATION · ANTHROPIC

Blameless postmortem, new controls

Continuous monitoring of eval transcripts for unexpected behavior
Rigorous assurance work with third-party eval vendors (Irregular)
METR independent review; redacted PyPI transcript within a week
Separation held: this is mitigation, not the incident, and not a product launch

Why it’s one segment, three boxes

Incident (Anthropic) ≠ context (OpenAI’s July 21 HF breach) ≠ product (Codex Security CLI). The narrative connection: evaluation containment failed at both labs within a fortnight while reusable control tooling shipped separately. Codex Security is not presented as the fix for either incident.

learn.chatgpt.com/docs/security/cli mlq.ai (client-only scope) anthropic.com (mitigations)

Weekly Claw #23

SEG 02 / ECONOMICS · 9.1

Segment 02 · OpenAI operational reset · part A: economics

Luna drops 80%. Not 90. Eighty.

Model	Input / M (old → new)	Output / M (old → new)
GPT-5.6 Luna	$1.00 → $0.20	$6.00 → $1.20
GPT-5.6 Terra	$2.50 → $2.00	$15.00 → $12.00
GPT-5.6 Sol	$5.00 (unchanged)	$30.00 (unchanged)

Source: OpenAI developer community pricing table + OpenAI pricing post, Jul 30, 2026. Sol adds Fast mode: 2.5× speed at 2× price.

Why so fast: GPT-5.6 helped optimize its own runtime — 20% lower serving cost (GPU kernels), >15% better token-generation efficiency (speculative decoding). Vendor-reported.
Auto-review in ChatGPT/Codex moves to Luna — OpenAI expects it to cost ~10× less.
Vendor framing, labeled: Luna ≈ year-ago frontier at “~6 cents on the dollar per task”; “~99% lower” cost per task than Fable 5 on Agents’ Last Exam.
Three weeks after launch. Price cuts usually take quarters.

The story isn’t that OpenAI got cheaper. It’s that efficiency gains are now big enough to reprice a product line in 21 days.

openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6 community.openai.com (price table) cnbc.com axios.com

Weekly Claw #23

SEG 02 / HARNESS · 9.1

Segment 02 · OpenAI operational reset · part B: harness-created capability

Two settings tripled a benchmark.

OpenAI re-ran ARC-AGI-3 with retained reasoning (pass previous_response_id) and compaction (summarize old context instead of rolling truncation at 175K chars).
Same model. Same tasks. Different plumbing: the model stopped re-deriving each game from scratch every turn.
OpenAI’s own conclusion: benchmarks measure harness design as much as the model.
Caveat stack: vendor-reported; public task set; not an independent rerun.

ARC-AGI-3 PUBLIC SET · VENDOR-REPORTED

ARC Prize verified setup

13.3%

OpenAI alternative harness

38.3%

Output tokens

~6× fewer

GPT-5.6 Sol (max), Relative Human Action Efficiency. Source: OpenAI research post.

If two API settings triple the score, last month’s leaderboard was measuring the plumbing.

openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores developers.openai.com (compaction docs)

Weekly Claw #23

SEG 03 / OPEN WAVE · 8.9

Segment 03 · the open language wave gets downloadable

Two agentic models you can actually download today.

DEEPSEEK V4 FLASH · OFFICIAL 0731

Post-training upgrade, same architecture

Jul 31: preview (Apr 24) graduates to official public beta — same 284B / 13B-active MoE, 1M context; only re-post-trained
Native Responses API + specific adaptation for Codex-style agents; config path backs up existing setup before writing provider metadata (reversible)
Weights on Hugging Face under MIT — downloadable now
V4 Pro unchanged; official Pro release “soon”; Codex support for Pro expected early August

THINKING MACHINES · INKLING-SMALL

A quarter of Inkling’s size, full weights

276B total / 12B active MoE; up to 1M context (64K/256K on Tinker); text/image/audio in, text out
Full weights on Hugging Face under Apache-2.0; NVFP4 quantized checkpoint alongside BF16
Fine-tunable on Tinker; vLLM/SGLang serving; community dual-DGX-Spark NVFP4 recipe exists
Benchmarks (SWE-bench Verified, Terminal-Bench, HLE, AA Index) are vendor-reported

The Hugging Face layer is the story: licenses, files, and quantizations you can verify — not launch-post adjectives.

huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 api-docs.deepseek.com/updates DeepSeek Codex docs thinkingmachines.ai/news/inkling-small huggingface.co/thinkingmachines/Inkling-Small

Weekly Claw #23

SEG 03 / THE TABLE · 8.9

Segment 03 · comparison · unknowns left blank, vendor claims labeled

The downloadable tier, side by side.

	DeepSeek V4 Flash (0731)	Inkling-Small	Kimi K3
Architecture	MoE, 284B total / 13B active	MoE, 276B total / 12B active	MoE, 2.8T total / ~104B active (KDA + AttnRes)
Context	1M in / 384K out*	1M (64K/256K on Tinker)	1M
Modalities	Text (tool-calling tuned)	Text + image + audio in → text	Text + vision
License	MIT	Apache-2.0	Kimi K3 License (custom)
Weights	Downloadable (HF)	Downloadable (HF, BF16 + NVFP4)	Downloadable (HF, MXFP4 ~1.4–1.56 TB)
Quant / local path	4-bit GGUF ~155 GB*	NVFP4 checkpoint; dual-Spark recipe (community)	Unsloth 1-bit GGUF 594 GB — needs ~610 GB+
API price / M	$0.14 in / $0.28 out*	unknown (Tinker metered)	unknown (platform.kimi.ai)
Benchmarks	Vendor-reported (agentic/coding up vs preview)	Vendor-reported (SWE-bench, Terminal-Bench, HLE)	Vendor-reported (“strongest open model”)
Caveat	Flash-only upgrade; Pro pending	Serving economics unverified	Self-host = multi-node; not one Mac

* Third-party-listed figure (dev.to decision guide / OpenRouter listing), not a primary DeepSeek price page. Blank/unknown cells are intentional — nothing invented.

HF: DeepSeek-V4-Flash-0731 HF: Inkling-Small HF: Kimi-K3 dev.to (*labeled figures)

Weekly Claw #23

SEG 04 / LOCAL REALITY · 8.9

Segment 04 · Kimi K3 · open weights ≠ locally runnable

“Full K3 on one 512 GB Mac?” Busted

MEMORY FLOOR · VERIFIED

Native MXFP4 weights

~1.4 TB

Unsloth 1-bit GGUF

594 GB

Unsloth guidance (RAM+VRAM)

~610 GB+

512 GB Mac Studio

falls short

Bar lengths illustrative, not to scale. Sizes from Unsloth docs & hardware analyses.

Kimi K3 weights (Jul 27): 2.8T-param open model, ~104B active, Kimi Delta Attention, native vision, 1M context — first open 3T-class model. Real and downloadable.
Full-model claim fails: smallest public unpruned quant is 594 GB; guidance is ~610 GB+ combined memory. One 512 GB Mac Studio is below that floor.
Honest paths: full model needs a cluster or accelerator supernode. Expert-pruned REAP80 fits (~350 GB, ~5.5 tok/s self-reported on M3 Ultra) but its card warns of noticeable quality degradation.
MoE sparsity cuts compute, not memory: all experts must stay addressable.

“Open weights” tells you what you may download. It says nothing about what you can run.

github.com/moonshotai/kimi-k3 unsloth.ai/docs/models/kimi-k3 HF: unsloth/Kimi-K3-GGUF modemguides.com (hardware check)

Weekly Claw #23

SEG 05 / MEDIA FRONTIER · 8.6

Segment 05 · three video-model launches · three product strategies

Three video models landed. The control layer is the fight.

MINIMAX H3 / HAILUO 3.0 · JUL 31

Omni-modal video, native stereo

2K, up to 15s, native stereo; text/image/video/audio context
Omni-reference: up to 9 images, 3 videos, 3 audio clips
Live via Hailuo + API; open weights promised, not downloadable at airtime

SEEDANCE 2.5 · BYTEDANCE / DREAMINA · JUL 31

Long takes plus targeted editing

Native 30s generation; long-video mode up to 3 minutes
Edit characters, environments, and camera moves with 1-second timestamp control
Up to 50 multimodal references; Dreamina subscribers in launch regions

FLUX 3 · BLACK FOREST LABS · JUL 23

One Self-Flow model: image, video, audio, action

Video up to 20s with native audio; gated early access
Image, video, audio, and action share one multimodal backbone
FLUX 3 Dev weights promised later in 2026; not downloadable at airtime

Same week. Different wedge.

MiniMax sells multimodal reference + audio. Seedance sells long takes + surgical editing. FLUX sells one backbone across media and action. The model is only half the product.

minimax.io/blog/minimax-h3 Dreamina: Seedance 2.5 official launch bfl.ai/blog/flux-3

Weekly Claw #23

SEG 05 / DEMOS · 8.6

Segment 05 · verified official demos · cue manually · no autoplay

Show, don’t describe.

MiniMax H3

Official 60-second launch film

Local · 60s

Full 60s filmOfficial @MiniMax_AINo autoplay

Launch film, not one raw outputOpen post ↗

Seedance 2.5

Official global launch showcase

Local · 2m40s

30s nativeTargeted editingNo autoplay

Official Dreamina launch assetOpen post ↗

FLUX 3 · Black Forest Labs

Official multimodal video demo

Remote · 145 MB

Frame from the official FLUX 3 video demo showing a running horse

Official demo frameVideo opens separatelyNo preload

Stream from BFL CDNPlay demo ↗

Weekly Claw #23

03 / SIGNAL FROM OUTSIDE · 7 MIN

Signal From Outside · weekly video review · permanent anchor · 7 min

Jensen Huang: “a very Linux moment.”

THIS WEEK’S VIDEO · Y COMBINATOR

“Jensen Huang: The Mindset That Built NVIDIA”

Garry Tan interviews Jensen Huang, live at Chase Center · YC Startup School 2026 · published July 26 · ~49 min. About twenty minutes in, Huang calls OpenClaw “a very Linux moment” and says NVIDIA told Peter Steinberger: “all of NVIDIA’s engineers are your engineers” — the same offer went to the Hermes team.

Official YouTube thumbnail (img.youtube.com) · fallback still · not a screenshot

THE STEERABILITY THESIS

Capability is solved enough. Steering is the frontier.

Agents don’t need to be 100% right — 80 or 99% is fine if humans can steer the remainder. Fine-grained controllability — change one word in a plan file, get a contained delta instead of a different output — is “probably the single biggest breakthrough we need for agents at every level.”

Jobs frame: AI eliminates tasks, not jobs — software employment +10% YoY, radiology +20% (Huang’s on-stage figures, not independently audited).

NO AUTOPLAY: the video is not embedded. If a moment is referenced on camera, Andy opens the official YouTube watch page manually in a browser tab; otherwise hold on the thumbnail card above.

youtube.com/watch?v=I4B37S1dyQQ (official · Y Combinator) ycrootaccess.com (full transcript)

Weekly Claw #23

04 / HOT TAKE · 4 MIN

Hot take · 4 min

“Open” now means three different things.

This week “open” shipped as downloadable MIT/Apache files (DeepSeek, Inkling-Small), as a custom-license 1.4 TB artifact you can’t run alone (Kimi K3), as a promise with a date TBD (MiniMax, FLUX) — and as a client whose engine stays closed (OpenAI’s security CLI). The word is doing four jobs.

Downloadable

DeepSeek V4 Flash (MIT), Inkling-Small (Apache-2.0). Verify the files, not the post.

Announced

MiniMax H3 (“coming days”), FLUX 3 Dev (“later 2026”). News, not a release.

Client-only

Codex Security CLI: Apache-2.0 wrapper around a gated cloud engine. Useful, but read the scope.

Prediction

By the end of Q3, “open weights” claims get audited like benchmark claims: no downloadable artifact with a license file, no headline.

Push back, Henry: Is a gated-engine open client still a win for the ecosystem — or openwashing with better PR?

Weekly Claw #23

SPONSOR

Also brought to you by

Trusted phone systems from trusted people. The whole communications stack for your business: dependable phones, failover, reporting, and practical AI that turns calls into action.

One accountable provider who actually answers.

heritagetel.com

Weekly Claw #23

05 / CLOSE

One to watch + where to follow

See you next Friday.

The transcripts

Anthropic promised a lightly redacted PyPI incident transcript within a week, and METR’s independent review is pending. Read both against OpenAI’s July 21 disclosure.

The weight drops

MiniMax H3 weights “in the coming days”; FLUX 3 Dev “later in 2026”; DeepSeek V4 Pro official “soon.” First downloadable artifact with a license file wins the headline.

The harness gap

Do other labs re-run their flagship benchmarks with retained reasoning + compaction — and how many leaderboard moves this quarter are plumbing, not models?

Follow the excitement.

Weekly Claw #23

SOURCES / LINKS

Every claim, one click · verified 2026-07-31

Sources & Links.

SEG 1 · SECURITY

anthropic.com/news/investigating-incidents-cybersecurity-evals cnbc.com/2026/07/30/anthropic-says-claude-gained-unauthorized-access cnn.com/2026/07/30/tech/anthropic-ai-models-break-out-hack bleepingcomputer.com (PyPI / 15 systems / 9,000 targets) cyberscoop.com/anthropic-claude-ai-hacks-real-companies learn.chatgpt.com/docs/security/cli mlq.ai (Codex Security CLI scope) theagenttimes.com (Apache-2.0 license)

SEG 2 · OPENAI RESET

openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6 community.openai.com (full price table) openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores developers.openai.com/api/docs/guides/compaction cnbc.com/2026/07/30/open-ai-price-cut-gpt axios.com/2026/07/30/openai-cuts-prices

SEG 3 · OPEN LANGUAGE WAVE

huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 api-docs.deepseek.com/updates api-docs.deepseek.com/quick_start/agent_integrations/codex thinkingmachines.ai/news/inkling-small huggingface.co/thinkingmachines/Inkling-Small NVIDIA forum (NVFP4 dual-Spark recipe, community) dev.to (third-party price/quant figures)

SEG 4 · KIMI K3

github.com/moonshotai/kimi-k3 huggingface.co/moonshotai/Kimi-K3 unsloth.ai/docs/models/kimi-k3 huggingface.co/unsloth/Kimi-K3-GGUF kingy.ai (Mac memory analysis) modemguides.com (hardware reality check) oflight.co.jp (self-host specs)

SEG 5 · MEDIA WAVE

minimax.io/blog/minimax-h3 platform.minimax.io/docs/guides/video-generation MiniMax H3 official 60s launch film Dreamina Seedance 2.5 official launch scmp.com (H3 license + weights timing) bfl.ai/blog/flux-3 FLUX 3 official demo mp4 (cdn.sanity.io, BFL CMS) globenewswire.com (BFL launch release) decrypt.co (FLUX 3 + FLUX-mimic)

SIGNAL FROM OUTSIDE

youtube.com — Jensen Huang: The Mindset That Built NVIDIA (YC Startup School 2026) ycrootaccess.com (full transcript)

BENCH / CAVEAT TEXTURE

arxiv.org/abs/2607.27080 (agent memory attack surface, bench) arxiv.org/abs/2607.27191 (research-judgment eval, merged caveat)

Vendor-reported claims are labeled on each slide. OpenAI pages block direct curl (bot protection); content was retrieved and cross-checked via independent reporting (CNBC, Axios) and the OpenAI developer community mirror.