OpenAI’s internal Astra solves ten open math and theoretical-CS problems for about $2,000; same day OpenAI cuts Luna and Terra pricing by 20–80%

On August 1, OpenAI co-founder Greg Brockman published a post announcing that an internal version of Astra — OpenAI’s next-generation flagship — had resolved ten open problems across mathematics and theoretical computer science for a combined Sol-API cost of roughly $2,000. The set covers existence of non-sofic groups, a counterexample to the Connes rigidity conjecture, advances in von Neumann algebras and high-dimensional sphere packing, and a new circuit-complexity lower bound; every result ships with a Lean formal certificate and a chain-of-thought derivation. On the same day, OpenAI cut Luna and Terra API prices by a further 20–80%, and Luna Max lines up against Sol Medium on Artificial Analysis’s Agentic Coding benchmark at roughly 1/25 of the latter’s price. Anthropic released Claude Opus 5 in parallel, scoring a perfect 42/42 on the IMO 2026 benchmark at about half the price of Fable 5. DeepSeek put out a GGUF-quantized build of V4 Flash 0731 that runs end-to-end on a 192 GB desktop without losing points. The Munich Regional Court ruled that Suno infringed copyright on both the training and output sides and rejected its fair-use defense, becoming the first adverse ruling against a music-generation model in an EU jurisdiction. These three storylines pushed “frontier capability, inference cost, and compliance boundary” to the center of industry attention on the same day.

Theme 1: OpenAI’s internal Astra solves ten open problems for about $2,000

The ten results Brockman listed span several long-standing threads in mathematics and theoretical computer science: proving the existence of non-sofic groups (which negates certain universality conjectures in group theory), producing a counterexample to the Connes Embedding / rigidity conjecture, advancing work on von Neumann algebras and high-dimensional sphere packing, and establishing a new circuit-complexity lower bound. OpenAI has published all ten proofs, each accompanied by a Lean formal certificate and a step-by-step chain-of-thought derivation; the combined Sol-API cost of running them is stated as roughly $2,000.

Mechanically, Astra is described by OpenAI as the next-generation multi-agent model family and is a different generation from the GPT-5.6 Sol / Luna / Terra lineup. OpenAI frames “ten independent breakthroughs on open problems in one run” as a first, and writing the cost as ~$2,000 is a deliberate choice to lock the implicit “brute-force search with massive trial and error” narrative to a controllable price tag. The team also publicly acknowledges that they tried other major open problems — including one from the Millennium Prize set — and failed, which is a candid concession that mathematical results do not automatically translate into broader scientific conclusions.

Why it matters: this is the first time a lab has produced ten results on long-standing open problems in formal domains — mathematics and theoretical computer science — under the model rather than the human-researcher identity, and each one carries a machine-checkable formal certificate. Reproducibility, the transfer of these methods to non-formal domains, and the training-inference coupling that produced them are all undisclosed. Yuchen Jin, working backward from Fable 5’s evaluation, called the problems “possibly Fields Medal-level”; Gary Marcus repeated his “strong math ≠ general intelligence” line and pointed out that at least one math PhD student (@henryquantum) has already found a clarity defect in one of the proofs. Sloan researcher poly Noamial (@polynoamial) also confirmed that they tried other major open problems, including a Millennium Prize item, without success.

Controversy clusters around three points. First, methodological opacity: OpenAI has not disclosed the success-to-attempt ratio across the ten problems, whether the system invoked formal tools such as Lean or Coq, or where human researchers intervened in the loop. Marcus lumps the reaction into the same category as the early community response to OpenAI’s first large models — results precede methods, and what the public sees is a 249-page results document rather than a mechanisms document. Second, the equivalence between formal proof and open-problem resolution: Tibo (@thsottiaux) framed the week as “efficiency gains, this weekend ten scientific breakthroughs”; Bojan Tunguz called it outright “another Deep Blue moment.” Third, whether this counts as AGI or ASI: Marcus anchored his argument in a 2024 public wager with Brundage — he would accept “5/10” on Astra as meeting his AGI bar, and the current public evidence is far below that line. For developers, the most direct consequence is that the test-time-compute engineering track now has its first reproducible script of “milestone results at a verifiable per-task cost.”

Sources:

Theme 2: OpenAI cuts Luna and Terra prices by 20–80%, and sol-advisor enters production usage

On the same day, OpenAI pushed the API price of Luna and Terra in the GPT-5.6 lineup down another notch — Luna by 80% and Terra by 20%. Developer communities immediately began using a hybrid pattern with Luna Max in front and Sol Medium in the back. The sol-advisor pattern promoted by Dan McAteer uses GPT-5.6 Sol as an orchestrator / advisor and dispatches Luna Max and Terra Max as implementers; the Artificial Analysis Agentic Coding scatterplot shared by @rasbt shows Luna Max landing near Sol Medium and Terra Max near Sol High, at a small fraction of the latter two’s prices.

Mechanically, Sol / Luna / Terra are different tiers within the same GPT-5.6 generation, corresponding to high, medium, and low reasoning intensity respectively. Combining the price cuts with intelligent routing means a single task can cost substantially less than the legacy “Sol end-to-end” pattern. Emad Mostaque estimated that, even with prompt caching factored in, a single Astra math case still sits at roughly 100 million cumulative tokens — still well below a week of Codex subscription quota.

Why it matters: model companies are no longer competing on benchmark scores alone; “cost per task in dollars” has moved onto the same axis. This rewrites the unit economics of agent applications — a run that used to cost one to three dollars per agent task can now drop to a few cents — and high-frequency AI services have, for the first time, a plausible path to scaled customer acquisition.

Sources:

Theme 3: DeepSeek V4 Flash 0731 ships, with 284B total / 13B active parameters fitting on a 192 GB desktop

Unsloth released a GGUF-quantized build of DeepSeek V4 Flash 0731: 3-bit at about 103 GB (runs on 110 GB total memory), 4-bit at about 155 GB, and Q8 lossless at about 162 GB. The model itself has 284B total / 13B active parameters, with about 96% of the routed-expert weights stored natively in FP4; the Q8 lossless build is therefore only about 7 GB larger than the 4-bit build, and Unsloth does not need to requantize. On benchmarks, V4 Flash 0731 reaches Terminal Bench 2.1 82.7, DeepSWE 54.4, and NL2Repo 54.2, beating the earlier V4 Pro Preview on multiple agent and coding tests.

Mechanically, DeepSeek built quantization-awareness into the MoE routing layer — FP4 is written into the weights themselves. When the community repackages the model into GGUF, the work is essentially repackaging rather than re-quantization, which preserves a layer of capability above V4 Pro Preview. Max For AI read this alongside Sam Altman’s “the GPT-5.4 flagship capability from four months ago is now available at 1/13 of the token price” and the Tsinghua-affiliated ModelBest “density law” (capability per parameter roughly doubling every three months), and the three threads point at the same conclusion: intelligence is rapidly becoming a commodity.

Why it matters: a model can run end-to-end on a single desktop workstation with 192 GB of unified memory, and the local build does not lose points. “Frontier-tier capability on a sub-few-tens-of-thousands-of-dollars local workstation” is starting to hold. Geekbb’s hands-on test of V4 Flash + Codex on the pelican SVG task took about 15 minutes 46 seconds and roughly 60,000 tokens per request; DeepSWE 54.4 means SWE-class tasks now have a viable all-local alternative. The Artificial Analysis cost-per-task numbers nail the cost axis down further: V4 Flash at a composite task cost of $0.03 with a composite index of 50; Claude Fable 5 at $3.15 with a composite index of 60 — roughly 105× cheaper at about 83% capability coverage.

A caveat: Lanshu’s independent testing found that V4 Flash’s front-end output did not clearly beat GLM-5.2, and the pure-HTML generation ranking on Arena sits lower than expected. DeepSeek’s official build has a maximum reasoning-depth option, and the version surfaced on Arena looks less rigorous than that. V4 Flash is therefore best treated as “fast first draft + low cost,” with stronger models still needed for fine editing. Teams such as Lobehub have already written “DeepSeek for the front end, GPT for hard tasks” hybrid routing as the default strategy in local agents like OpenCodex; Indie Fox’s independent testing confirms that the cost / quality / speed trade-off holds across all three dimensions.

Sources:

Theme 4: Anthropic ships Claude Opus 5 — a 42/42 on IMO 2026 at less than half of Fable 5’s price

AI Valley reposted Anthropic’s official announcement: Claude Opus 5 launched simultaneously on apps, Claude Code, and the API, positioned as “approaching Fable 5 intelligence at roughly half the price.” Headline numbers: a perfect 42/42 on the IMO 2026 benchmark, simultaneous state-of-the-art results across coding, reasoning, agentic search, and computer use, and an explicit safety emphasis layered on top.

Mechanically, the price / safety / capability triangle is being pulled tight at the same time, which signals that Anthropic’s commercial strategy has shifted from “matching OpenAI’s strongest model” to “matching the best price-performance point.” Cursor launched a smart AI Router the same day, making the strategy a product — automatically picking among tiers such as GPT-5.6, Claude Sonnet 5, and Claude Fable 5 according to task characteristics.

Why it matters: the Claude Opus 5 + Cursor Router combination rewrites the market narrative from “frontier model = single flagship” to “intelligent routing per subtask.” For developers, the most direct implication is to plan the stack around “what I can afford” rather than “the best model I can reach.”

Sources:

https://www.theaivalley.com/p/anthropic-s-opus-5-is-almost-fable-5

Theme 5: YC open-sources the QM multi-agent harness, with the tool table grown past 350 entries and Slack / Web native

YC, with Garry Tan, publicly released QM Harness, the internal multi-agent system it has been using for some time. QM scopes memory, files, keychain, permissions, cron, and sandbox on a per-user / per-channel basis; both the harness and the model are replaceable, with Pi, OpenCodex, Codex, and Claude Code all sharing the same kernel. It supports Slack and Web natively, ships under the MIT license, and can be deployed to Fly or AWS with a single command. YC’s internal tool registry has grown from an initial ~20 entries to more than 350, covering accounting, legal, event operations, and engineering collaboration — real business workflows rather than demos.

Mechanically, QM open-sources the entire engineering abstraction “Agent = model + sandbox + tool table + memory + permissions” and explicitly decouples the harness from the model. This resonates strongly with a parallel line of activity in the industry (OpenCodex, Memmy, and other local agents) — QM is the platform play, the local agents are the developer-side decoupling play.

Why it matters: the industry judgment that “agent shells and models are decoupling” gets its first open-source implementation, already validated against 350+ tools and deployable. SMBs can now run multi-agent workflows on Slack and Web the way YC does, without being locked to a single model vendor. Anna Zhang noted that YC already broadcasts all agent conversations internally by default, effectively making “organizational memory” a default artifact — a style consistent with the AGENTS.md / CLAUDE.md pattern OMARSAR recommends, but packaged into a deployable project. Lucas Beyer described it as “an OpenClaw-for-work optimization,” directly usable by small teams that do not want to build a harness from scratch.

Sources:

Theme 6: Munich Regional Court finds Suno liable on both training and output, and rejects the fair-use defense

The Munich Regional Court ruled on August 1 that the AI music generator Suno infringes copyright on both the training side and the generation side, and rejected its fair-use defense. The court found that Suno versions 3.5 and 4 can reproduce original elements from six well-known songs — a finding of “memorization” infringement — and assigned liability to Suno rather than to users of Suno. The ruling also clarified that US fair-use doctrine does not apply in this case. The judgment is not yet final, but it is the first adverse substantive ruling against a generative music model in an EU jurisdiction.

Mechanically, the court separated the question “what is reproduced in the model weights” from “what is reproduced in the model’s outputs,” rejecting both the “training-side fair use” defense and the “output-side pure-tool exemption” defense in one move. If the ruling becomes final, EU music models may be forced to introduce training-data source auditing and output deduplication.

Why it matters: this is the first formal finding in an EU jurisdiction of end-to-end liability for a music model based on “memorization infringement.” It sets a reference line for subsequent copyright decisions on image, video, and code generative models, and is likely to influence the EU AI Act’s enforcement of training-data transparency clauses. The judgment is not yet final; Suno can still appeal, and it would be premature to extrapolate this into a settled rule.

Sources:

https://the-decoder.com/german-court-rules-ai-music-generator-suno-violated-copyrights-rejects-fair-use-defense

Theme 7: Huawei open-sources openPangu-2.0-Pro — the first 500B+ frontier model trained entirely on non-NVIDIA hardware

Huawei officially announced the open-source release of openPangu-2.0-Pro: 505B parameters, MoE architecture, and described as the first 500B+ frontier model trained end-to-end on non-NVIDIA hardware (Ascend NPUs). meng shao reposted YC’s internal QM Harness open-source announcement in parallel.

Mechanically, openPangu-2.0-Pro sits at a distance from the global mainstream frontier stack at the hardware layer. Read alongside OpenAI, DeepSeek, and Moonshot Kimi K3 all pushing model costs down at the same time, the global frontier-model landscape is, for the first time, splitting into two tracks: US vendors dominate the open-source ecosystem, and Chinese vendors dominate the non-NVIDIA compute stack.

Why it matters: when a 505B-scale model can be trained entirely off the NVIDIA stack, it proves that non-CUDA training engineering is now production-ready. For China’s compute supply chain, training frameworks, and the domestic NPU ecosystem, this is one of the most substantive developments of 2026 so far.

Sources:

https://x.com/shao__meng/status/2083520825601434044

Theme 8: Snapchat blocks fully AI-generated videos from the Spotlight recommendation feed

Starting August 1, Snap formally rejects any “entirely AI-generated” video from the Spotlight recommendation pool. The rule is not “down-rank AI-labeled content” — it removes recommendation eligibility outright. Videos in which AI is used as an aid for editing, enhancement, or voice-over can still be recommended, but must be explicitly labeled as such. Max For AI read this alongside YouTube tightening monetization eligibility for templated repetitive content and LinkedIn / Substack handling AI slop.

Mechanically, platforms are stepping up from “content distributor” to “gatekeeper of content authenticity.” Fully automated generation is now explicitly separated from “AI-assisted + human-authored,” and AI slop has shifted from “labelable content” to “throttled content.”

Why it matters: this is the first time a content platform has, at the product-rule level, separated the punishment of “AI replacing the creator” from “AI assisting the creator.” It is a direct blow to pure-AI short-video matrices and bulk content farms, and a tailwind for creators who depend on AI tools but maintain a personal voice. The same day, Microsoft AI VP Nando de Freitas announced his departure (he had been recruited by Mustafa Suleyman in September 2024); he said Microsoft AI already has a complete model lineup (MAI-Image-2.5, MAI-Voice-2, MAI-Thinking-1) and the next step is “continuous model iteration” — which is precisely the moment he chose to leave.

Sources:

Theme 9: Figure F.03 climbs a ladder fully autonomously — humanoid robots clear a “human-built infrastructure” bar

Figure released a video of robot F.03 climbing a real metal ladder step by step and standing steadily on the platform, with no remote control and no cloud fallback. The Stanford AI Lab cited Annie Chen’s research to note that running real-time interactivity, task understanding, and full-body action prediction simultaneously is one of the few remaining hard problems for humanoid robots.

Mechanically, a ladder is the canonical “built for humans, unusable by robots” facility. F.03’s visual perception, hand-foot coordination, and real-time center-of-mass control all close the loop locally, which means it no longer demands that the world reshape itself around it.

Why it matters: the humanoid field today is dominated by Unitree’s backflips and dance demos — “demo-grade” actions. Climbing a ladder is zero difficulty for a human but a 0-to-1 milestone for a humanoid; it is the entry bar for “non-flat-ground” settings such as construction sites, mezzanine maintenance, and elevated work. The climb significantly shortens the distance between “lab demo” and “real job site.”

In the short term, the cost and reliability of this capability are still limited: AYi described F.03 himself as “moving like an 80-year-old man” — slow and cautious, with each rung requiring re-localization. “Can do it autonomously” and “does it fast” are two independent engineering curves, however; the former is the 0-to-1, the latter is just iteration on existing algorithms. Read alongside the other threads released on the same day — Snap blocking fully AI-generated videos, Suno’s adverse EU ruling, OpenAI cutting model costs another 20–80% — this can be read as the starting point for “embodied intelligence enters the industrial deployment list” in the second half of 2026: hardware that closes the loop, businesses that can do the math on cost, and regulators that can accommodate it, all holding at the same time for the first time.

Sources:

Theme 10: Karpathy’s autoresearch idea ported to Claude Code / Codex, with overnight agent self-iteration becoming a product

GitHubDaily’s recommended open-source project, Autoresearch, takes Karpathy’s “let the model run a hundred experiments overnight” idea and turns it into a 14-subcommand Claude Code / Codex skill: given a target and a quantitative metric, the agent loops on editing code, running validation, keeping what improves and rolling back what does not. Nine safety hooks ship in by default to prevent destructive operations.

Mechanically, this is “the agent writing the agent” packaged as a repeatable workflow rather than a one-off script. For individual developers, “set a goal, let it run overnight, check the result in the morning” is becoming the new normal.

Why it matters: once an agent can iterate on itself overnight, the human-in-the-loop cycles that model evaluation, skill factories, and A/B experiments used to require can all be machine-driven. Read alongside the test-time-compute bet both OpenAI and Anthropic are making, this is one of the most important engineering paradigms of the second half of 2026.

Sources:

https://x.com/GitHub_Daily/status/2083492943521010021

High-value briefs

Nando de Freitas leaves Microsoft AI: announced August 1; previously an Oxford professor and DeepMind chief scientist for nearly a decade, and at Microsoft for less than two years. He said the Microsoft AI team is already mature, but the in-house model lineup has just been assembled and a core multimodal lead is choosing to leave — the timing is telling. Next step undisclosed; he is headed to the Argentine Andes on vacation first.
DeepSeek Harness opens closed beta: DeepSeek Harness lead tianyi publicly recruited beta testers on X, with the stated requirement being prior work on agent-harness open-source projects; commenters including @lobehub, Pi, and OpenCodex responded. The DeepSeek model + official harness pairing is being treated as the next key land grab in agent infrastructure.
Cursor smart AI Router launches: dynamically routes each task among tiers such as GPT-5.6, Claude Sonnet 5, and Fable 5, matching Anthropic Opus 5’s “roughly half of Fable’s price” positioning.
Yann LeCun and Zuckerberg refollow each other: Yann LeCun’s X account was refollowed by finkd, which Chinese-language commentators are reading as a possible signal of Meta’s internal posture toward the frontier-model track shifting.
AI spread vs. AI cost polarization: Sam Altman said “the GPT-5.4 flagship capability from four months ago is now available at about 1/13 of the token price”; Max For AI noted in parallel that “V4 Flash is enough for first-draft front-end work, only escalation to a stronger model for fine editing” — “route by task” on the product side is no longer optional.
Pfizer’s Graph RAG production case: against Gartner’s prediction that 30% of generative-AI projects will fail, Pfizer’s team used a chunk-by-chunk graph-ingestion + parallel vector-store query architecture and compressed data-cleanup time from three months to three weeks.

🕐 Selected hourly signals

PT time	Signal	Why it is worth remembering
00:00	Astra publishes ten proofs of open math / theoretical-CS problems across multiple areas	Highest single-day density of open-problem breakthroughs by one model
00:30	Sol / Luna / Terra prices cut 20–80%; sol-advisor enters production usage	Beyond benchmark scores, “cost per task” becomes the main axis
02:00	DeepSeek V4 Flash 0731 GGUF quantization release	V4-tier runs on a 192 GB desktop; Q8 lossless is only 7 GB larger than 4-bit
04:00	Snapchat fully blocks AI-generated videos from Spotlight	A platform separates “AI replacing the creator” from “AI assisting the creator” at the rule layer for the first time
05:00	DeepSeek Harness opens closed beta	The DeepSeek model + official harness pairing begins to take shape
06:00	Anthropic Claude Opus 5 launches; perfect 42/42 on IMO 2026	The top-tier model price / capability curve gets pushed down another notch
07:00	YC open-sources QM Harness (350+ tools, Slack / Web native)	A multi-agent engineering abstraction becomes open source in a deployable form for the first time
08:00	Figure F.03 climbs a ladder fully autonomously	A humanoid robot closes the loop autonomously on human-built infrastructure for the first time
10:00	Microsoft AI VP Nando de Freitas announces his departure	In-house model lineup just completed; a key multimodal lead exits
11:00	Karpathy’s autoresearch idea lands in Claude Code / Codex	Overnight agent self-iteration moves from concept to ready-made skill

Editorial conclusion

August 1 was pushed to the center by several storylines at once: OpenAI’s internal Astra pushed the “model + formal proof + verifiable cost” track to a level never seen before; OpenAI simultaneously cut Luna / Terra prices by 20–80%, putting “cost per task” on the same axis as benchmark scores; and the Munich Regional Court dealt Suno a heavy blow and drew the EU compliance boundary for music models for the first time. Alongside these three, DeepSeek V4 Flash moved onto a desktop workstation, Anthropic Opus 5 hit a perfect score on IMO 2026, YC open-sourced the 350+ tool QM Harness, Snap fully blocked AI-generated videos from its recommendation feed, and Figure F.03 climbed a ladder fully autonomously — scattered on the surface, but pointing the same direction. Frontier capability, inference cost, content authenticity, and robot physical autonomy have each entered a usable inflection zone.

Sources and method

This daily is built by combining one morning AIHOT digest, two named sources (AI Valley and HubToday), and nineteen hourly X snapshot captures from the 2026-08-01-pt folder. All numbers, model names, prices, and benchmark scores are taken verbatim from the captures and have not been cross-checked against external documentation; vendor-reported benchmarks from OpenAI, Anthropic, and DeepSeek are cited as such. Treat any cost data as needing verification against the vendor’s own documentation before reuse.