Claude Fable 5.2 lowers prices to $8/$40 per million input/output tokens, but in the GDPx benchmark of professional tasks, cost per task rises from $67.16 to $88.15, as more tests and subagent checks offset the lower rates. Anthropic has launched its IPO roadshow, proposing to offer 825 million shares at an expected price of $70–$78 each; the registration statement is not yet effective. Google DeepMind’s AlphaProof Code generated more than 6 million lines of Lean proofs in four weeks, verifying specified security properties of the actual code implementing Signal’s core protocols.
Cheaper tokens do not necessarily mean cheaper tasks: execution volume and how work is organized also shape the bill.
Anthropic released Claude Fable 5.2, lowering prices to $8/$40 per million input/output tokens, 20% below 5.1. Yet in SemiAnalysis’s GDPx benchmark of professional tasks, cost per task rose from $67.16 to $88.15. Execution logs show that 5.2 sometimes writes hundreds to thousands of tests and launches subagents to perform independent checks, while 5.1 typically tries around 20 test cases.
Read original →Meta released Muse Ember for tool use, coding subtasks, and long-document parsing, with a 1-million-token context window and up to 512,000 output tokens. It is the first model trained using Muse Wonder Full: Muse Wonder handles planning, while Ember takes on large volumes of repetitive execution work. It is now available in Muse Code and through the Meta Model API.
Read original →Tencent released DSpine, a five-layer speculative decoding draft model for Qwen3.8-27B with a block size of 16. It generates draft tokens in parallel while preserving causal dependencies by passing features from preceding predictions between network layers. Tencent reports throughput of about 2.9–3.8× that of autoregressive decoding. Tests used SGLang 0.5.22, BF16, and FlashAttention 3, with concurrency set to 8 and ReplaySSM enabled.
Read original →A research team introduced s3, which uses cooperative reinforcement learning to let agents decide how long to work independently, when to share findings, and when to read teammates’ messages. Unlike pre-orchestrated multi-agent systems, it assigns no fixed roles; complementary divisions of labor emerge during training and transfer to larger teams. In experiments with matched training and inference budgets, doubling the team size relative to training yielded more than 25× as many solved coordination problems as the strongest baseline.
Read original →🔍 Analysis: With Fable 5.2 running more tests and subagent checks, a 20% token price cut still did not lower its cost per GDPx task. Muse Ember separates planning from repetitive execution, offering another way to organize execution workloads. Meanwhile, s3 learns how to divide work and when to communicate during training. In budget-matched experiments, it solved more than 25 times as many coordination problems as the strongest baseline, but that does not translate directly into a cost advantage on production tasks. DSpine’s reported 2.9–3.8× throughput gain applies to decoding and likewise cannot directly offset the additional execution steps agents take. For developers, comparing options requires accounting for unit prices, execution volume, and task outcomes—not just the price per million tokens.
The proposed price range is out, with proceeds from company-issued shares and existing shareholders’ sales going to different recipients.
Anthropic has launched its IPO roadshow, proposing to offer 825 million shares of Class A common stock at an expected price of $70–$78 per share, and has applied to list on Nasdaq under the ticker ANT. The offering includes shares sold by the company and by existing shareholders; the company will not receive proceeds from the latter. Underwriters will also receive a 30-day option to purchase additional shares. The related registration statement has been filed with the U.S. Securities and Exchange Commission but is not yet effective. These securities may not be sold, nor may offers to buy be accepted, before it becomes effective.
Read original →Separate efforts involving Signal and CPython connect Lean checks to actual code.
Google DeepMind introduced AlphaProof Code, whose agents generated more than 6 million lines of Lean proofs in four weeks to verify specified security properties of Signal’s core protocol code. Unlike protocol verification that largely stays at the design level, this effort reaches into the actual compiled code, covering message protection before and after a temporary device compromise. All proof goals were reviewed by experts. The system also flagged about 1% of the code for human review; broadening its applicability will require further research.
Read original →A project used Opus 5.5 to formalize CPython, generating 613,318 lines of Lean code that map Python’s internal implementation line by line. The code was tested against the actual interpreter and passed Lean checks. The project also uncovered real Python bugs, including crashes that can be triggered using only ordinary Python code.
Read original →🔍 Analysis: AlphaProof Code and the Opus 5.5 project both move verification closer to real implementations rather than abstract designs. The former proves specified security properties of Signal’s compiled code; the latter maps CPython’s internals line by line and tests the result against the actual interpreter. Their verification scopes should not be conflated: 613,318 lines of code passing Lean checks does not mean every security property of CPython has been proved. For engineering teams, the key measure of such results is not the number of lines of proof code, but whether the proof goals cover real risks and how the formal representation corresponds to the running code.
Apollo 2 achieves a 96% success rate sorting and placing automotive parts into trays, as the model rolls out to robotics partners.
Google released the vision-language-action model Gemini Robotics 2.5, testing its capabilities across more than 200 robotics tasks. Compared with previous baselines, task success rates improved by 30%–100%, and completion was 20%–30% faster. In a proof of concept on a Mercedes-Benz production line, Apptronik’s Apollo 2 robot sorted automotive parts and placed them into trays with a 96% success rate. The model is rolling out to robotics partners.
Read original →Stability across 100 editing rounds and native RGBA output address detail preservation and transparent asset generation, respectively.
ByteDance released the image model Seedream 5.5, emphasizing more realistic images and claiming improved stability across 100 consecutive editing rounds. To address detail loss from repeated revisions, the new model emphasizes pixel-level preservation of unedited regions, reducing detail degradation across multiple edits. It also supports native RGBA output, enabling direct generation of transparent and semitransparent assets.
Read original →