Sakana's Fugu models, Nvidia's Anthropic IPO talks, Senate AI safety rules, and OpenAI's Habitat
A concise daily brief on Sakana AI's Pareto orchestration models, Nvidia's anchor talks in Anthropic's $100B IPO, U.S. Senate bipartisan safety legislation, OpenAI's billion-user Habitat storage architecture, and Devin's automated testing with Astra.
This brief covers AI developments published between September 11 at 08:15 and September 12 at 08:15 in Dhaka. Over the past 24 hours, the industry balanced frontier model orchestration with the institutional and physical systems required to run them at scale. Primary developments include two new orchestration models from Sakana AI, confidential talks for Nvidia to anchor Anthropic's planned $100 billion initial public offering, bipartisan U.S. Senate negotiations on mandatory safety standards, OpenAI's engineering blueprint for storage serving one billion users, and Cognition's deployment of GPT-6 Astra for automated software verification.
Quick scan
| Development | What changed | Why it matters | What to watch |
|---|---|---|---|
| Sakana AI Fugu models | Sakana AI released Fugu Max for cost-efficient routing and Fugu Ultra v2 for peak reasoning, orchestrating open and specialized models including NVIDIA Nemotron. 1 | Orchestration provides frontier-level performance at lower token costs while removing dependencies on proprietary model providers. | Benchmark validation on SWEFish and independent latency comparisons via the OpenAI-compatible API. |
| Nvidia & Anthropic IPO | Reuters reported that Anthropic is in discussions with Nvidia to commit up to $10 billion as an anchor investor in an IPO seeking up to $100 billion at a $2 trillion valuation. 2 | Anchor commitments from primary chip suppliers solidify capital loops and test public market appetite for frontier AI labs. | Formal S-1 prospectus disclosures and antitrust regulatory assessments before the November elections. |
| Senate AI safety draft | Bipartisan Senate negotiators are drafting legislation to impose a legal "duty of care" on frontier AI developers to prevent catastrophic risks. 3 | Federal standards would replace voluntary commitments with judicial enforcement and could preempt conflicting state laws. | Final statutory text before the congressional pre-election recess and national laboratory evaluation requirements. |
| OpenAI Habitat storage | OpenAI published architecture details for Habitat, its application storage layer handling 70 million requests per second and 500 petabytes across 40 regions, alongside a migration to Rust. 4 | The writeup demonstrates how database proxying and model-assisted code rewrites keep tail latencies stable at global consumer scale. | Publication of Part II covering Azure Cosmos DB optimizations and partition strategies. |
| Cognition Devin verification | Cognition integrated GPT-6 Astra into Devin's testing harness to run code, record simulator playback, and verify fixes from customer bug reports. 5 | Autonomous testing moves agents from speculative code generation toward demonstrable verification before human review. | Long-horizon pass rates on multi-file pull request benchmarks. |
Sakana AI orchestrates the Pareto frontier with Fugu Max and Ultra v2
Tokyo-based Sakana AI launched two new systems in its Fugu orchestration family: Fugu Max and Fugu Ultra v2. 1 The company designed both systems around a shared routing architecture that distributes subtasks across a pool of open-weights and specialized models, including NVIDIA Nemotron models. 1
Fugu Max focuses on cost efficiency. By routing requests dynamically to the smallest model capable of solving them, Fugu Max sets pricing at $2 per million input tokens and $6 per million output tokens. 1 That output price is 40% to 60% lower than rates for Claude Sonnet 5, GPT-5.6 Terra, and Kimi K3. 1 In benchmark testing, Fugu Max recorded top scores across six evaluations, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish, Sakana AI's internal coding suite. 1
Fugu Ultra v2 targets maximum reasoning depth on complex tasks without incorporating proprietary frontier models into its agent pool. 1 Sakana AI excluded Claude Fable 5, Claude Fable 5.1, and GPT-6 Astra from the underlying model set. 1 On Chartography, an evaluation measuring visual reasoning and data interpretation, Fugu Ultra v2 scored 48.3, compared to 29.5 for Fable 5 and 27.3 for Opus 5. 1 On DeepSWE, a real-world software engineering benchmark, Fugu Ultra v2 achieved 74.3. 1 Both models are available immediately through Sakana AI's OpenAI-compatible API. 1
What to watch: Independent benchmark evaluations on SWEFish and API response latency under sustained multi-agent task loads.
Nvidia in talks for $10 billion anchor role in Anthropic IPO
Anthropic is in discussions to bring Nvidia onboard as an anchor investor for its upcoming initial public offering, according to reporting by Reuters. 2 Anthropic aims to raise up to $100 billion in the offering, targeting a total valuation near $2 trillion before the U.S. midterm elections in November. 2
Nvidia is evaluating an investment of up to $10 billion to serve as an anchor backer. 2 Anchor commitments allow large institutional backers to take designated allocations before public marketing begins, stabilizing demand for oversized listings. 2 The prospective agreement expands an existing partnership established in November 2025, when Nvidia pledged up to $10 billion alongside Anthropic's commitment to purchase $30 billion of Azure compute powered by Nvidia processors. 2
Anthropic's annualized revenue run rate exceeded $65 billion by late July, climbing from $9 billion at the end of 2025. 2 The startup continues to balance compute arrangements across multiple providers, including a $100 billion, 10-year agreement with Amazon for AWS infrastructure and one million Trainium2 chips, alongside Google TPU agreements and an internal custom silicon team. 2
What to watch: The filing of Anthropic's formal registration statement and confirmation of anchor allocations in public prospectus disclosures.
Senate negotiators consider federal duty of care for frontier AI developers
Members of the U.S. Senate are negotiating bipartisan legislation to establish a statutory "duty of care" for frontier AI model developers, according to two Senate aides and a lobbyist familiar with the discussions. 3 Senate Majority Leader John Thune, Senate Commerce Committee Chairman Ted Cruz, Senator Amy Klobuchar, and Senator Maria Cantwell are leading the talks. 3
Under the framework, developers would bear a legal obligation to design frontier systems to prevent catastrophic risks, focusing on biological, chemical, radiological, nuclear, and advanced cyber capabilities. 3 The proposal would empower the federal government to block releases of systems deemed unsafe, while granting developers the right to contest those administrative decisions in federal court. 3 Negotiators are also considering provisions to preempt state-level AI safety laws, establishing a single federal standard. 3
The legislative effort follows heightened scrutiny around autonomous agent breakouts and public warnings from technical researchers. 3 Senator Cantwell stated on X that meaningful legislation must mandate pre-deployment testing by scientists at U.S. national laboratories to evaluate weaponization and cyber risk. 3
What to watch: The release of formal draft text before the pre-election congressional recess and the division of authority between the Department of Commerce and federal courts.
OpenAI details Habitat storage platform for one billion users and Rust rewrite
OpenAI published an architectural breakdown of Habitat, its distributed online application storage platform. 4 Habitat now handles more than 70 million requests per second, supports products used by over one billion people each week, and serves more than 500 petabytes of data across nearly 40 geographic regions. 4
Habitat began in mid-2024 as a Python client-side library connected to Azure Cosmos DB, abstracting schema lookup, routing, authorization, and connection pooling for product developers. 4 In mid-2025, OpenAI transitioned Habitat into a centralized standalone service to establish unified control over deployments and access control policies. 4
To scale Python under heavy workloads, engineers tracked asyncio event loop delays, replaced LIFO connection reuse with FIFO to eliminate runaway load imbalances, and used Envoy proxies to multiplex HTTP/1 connections into HTTP/2. 4 Habitat uses a constrained NoSQL object-and-edge model inspired by Facebook's TAO, intentionally avoiding complex relational joins on online paths and offloading analytical queries to Rockset via change data capture streams. 4
In the second quarter of 2026, two engineers rewrote the entire Habitat service in Rust using Codex and GPT-5.5. 4 The Rust implementation now processes 95% of production traffic, demonstrating six times higher CPU efficiency and 15 times higher memory efficiency than the Python service, with full deprecation of the Python stack planned in the coming weeks. 4
What to watch: Part II of OpenAI's storage series, focusing on database-level partitioning and multi-region Azure Cosmos DB availability.
Cognition deploys GPT-6 Astra for autonomous verification in Devin
Cognition integrated OpenAI's GPT-6 Astra into the testing harness for Devin, its autonomous software engineering agent. 5 The deployment spans Devin's cloud agent, command-line interface, and desktop applications, focusing on verifying generated code before human review. 5
In one testing workflow, Devin uses Astra to test an iPhone game inside an emulator, returning a video recording of runtime behavior alongside a structured test report detailing passed checks and uncovered paths. 5 For customer support operations, engineers supply Devin with user-submitted bug screenshots; Astra reproduces the defect, implements the code fix, and returns an updated screenshot showing the resolved state. 5
Cognition co-founder Walden Yan stated that generating concrete verification artifacts allows engineering teams to reduce manual code inspection and accelerate software releases. 5
What to watch: Measurement of engineer review latency and bug recurrence rates across commercial Devin deployments utilizing automated visual testing.
What to watch
- Multi-model routing economics: Sakana AI's Fugu Max tests whether dynamic orchestration of open-weights models can sustain a 40% to 60% price advantage over closed frontier APIs. 1
- IPO anchor disclosures: Progress in Nvidia's talks with Anthropic will indicate whether frontier labs can secure strategic hardware capital ahead of historic public listings. 2
- Statutory AI boundaries: The bipartisan Senate duty-of-care proposal will test whether Congress can pass binding safety rules with national lab evaluations before midterm elections. 3
- Model-assisted infrastructure rewrites: OpenAI's Rust rewrite of Habitat offers a template for using coding agents to replace high-concurrency Python services. 4
- Autonomous software verification: Cognition's testing harness will reveal whether simulation recordings and visual proofs can reduce manual pull-request reviews. 5
References
- 1
- 2
- 3
- 4
- 5
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.
Related content
More from this channel
- Astra for Law, Claude's 26% of Anthropic's R&D, and a Cell cover won by a 9B model
- DeepMind's AGI institute, Vera Rubin's MLPerf debut, and OpenAI's 166-page proof
- Z.AI's $5B raise, Anthropic's Nasdaq choice, Trump's slowdown pushback, and Astra at Perplexity
- Amodei's call to pace the frontier, OpenAI's 2026 IPO delay, Meta's AI reorg, and Hyundai's Nvidia bet
- NASA-IBM's lunar model, OpenAI's Agents API, Anthropic's threat report, and Gemini for Windows
- Google's €13B nuclear pact, MAPL-EMIT from orbit, OpenAI's safety mandate, and Suno v6
- Muse, Mistral's €3B, Cognition's $48B, and AI at the lab bench
- Rentosertib's six aging clocks point to a biomarker signal