Search “Model releases”

Replit Launches Replit for Business, Proposing a Blueprint for the Self-Driving Company

Replit has released Replit for Business, offering enterprises a foundation that lets agents use company knowledge and systems within controlled boundaries, with the goal of helping enterprises transform into self-driving companies. Replit says that in the first half of 2026 its engineers' code output nearly tripled, while review time, rollbacks, and product incidents remained flat. Customer SMFL says a development process that originally required 6 people and 6 months was completed by 1 person in 1 week, and full lifecycle costs dropped from about 3 billion yen to under 1 billion yen.

Podcast: Kimi K3.1 Pro, the new cybersecurity reality, and the Muse user experience

In episode 20 of the Interconnects podcast, Nathan Lambert and Florian Brand discussed how Kimi K3.1 Pro ranks high on Artificial Analysis but receives mixed reviews, arguing that its training leans more toward exploitation than exploration, with weaker performance on long-horizon benchmarks such as ProgramBench. The two also discussed the cybersecurity controversy surrounding open-weight models, noting that no large-scale exploitation occurred after the disclosure of the PHP2Shell vulnerability, and talked about Anthropic not hiring junior employees as well as Florian's experience using Muse to handle everyday tasks such as shopping.

Behind the Kimi K3.1 Pro quota controversy: the stronger the model, the tighter the compute

Moonshot AI released Kimi K3.1 Pro on October 24, scoring 65.7% on Terminal-Bench 4.0, but users of the 199-yuan and 699-yuan plans complain that quotas are insufficient, and some developers estimate that the API value of the old plans is about 1.94 times that of the new plans. K3.1 Pro's per-million-Token input and output prices dropped from K3's 20 yuan and 100 yuan to 13 yuan and 68 yuan, while complex Agent tasks can consume millions of Tokens. A CAICT report shows that in June 2026, domestic daily average Tokens were about 500 trillion, nearly four times the figure in December 2025; the author believes that the compute shortage will still constrain independent model companies over the long term.

Moonshot AI releases Kimi K3.1 Pro: 2.8T total parameters, 104B activated parameters, weights scheduled to be opened on November 4

Moonshot AI releases Kimi K3.1 Pro, a multimodal MoE model with 2.8T total parameters, 104B activated parameters, and a 1 million token context window. The Standard and Highspeed versions are now available on the Kimi website, App, Kimi Work, Kimi Code and API. The company says Highspeed offers 2–3x faster output at twice the price of the Standard version. Artificial Analysis ranks it fourth on the Intelligence Index with a score of 55, behind Opus 5.5's 58; on Arena's Code Arena WebDev it ranks 8th with 1692 points, up nine places from K3. Standard pricing is $2 per million tokens for input, $10 for output, and $0.2 for cached input, one-third lower than K3; the weights have not yet been released, and the company promises to open them on November 4.

CAICT Report: In 2026, Average Prices of Domestic Large Models Doubled, and Per-Second Throughput Also Doubled

The "Large Model Token Service Performance Monitoring Report (2026)," jointly released by the China Academy of Information and Communications Technology (CAICT) and other institutions, shows that in August 2026, the average TPS of domestic models reached 78.23, double the 2025 average of 38.25; however, typical foreign models stood at 112.82 in the same period, still higher than domestic models. The normalized average price of domestic token services rose from 6.8 yuan in the second half of 2025 to 13.46 yuan, an increase of nearly 1x; in June 2026, average daily domestic token consumption was about 500 trillion, nearly 4 times that of December 2025. Among the monitored models, those with a context length of 1M accounted for 47.8%, and those with 200K and above totaled 95.7%.

Databricks on AI coding cost governance: unit costs cut by up to 90%

Patrick, a product lead at Databricks, reviewed the cost overrun problem after the company fully adopted AI coding agents, saying that by migrating to cheaper models, routing, spend visibility, and cutting token overhead, unit costs in some scenarios fell by as much as 90%. He noted that hard budget caps would interrupt the best developers, so the company switched to gradual friction through warnings, approvals, and model downgrades, and productized its internal solution as Unity Gateway. The team built its own evaluation using real PRs from its own codebase, finding that GLM 5.2 matched Opus 4.8 in quality at about one-third lower cost per task, while a model that was 1.7 times cheaper per token ended up costing more per task because it consumed more tokens.

Claude Platform now generally available on Google Cloud, operated by Anthropic

Claude Models on Google Cloud (Anthropic-hosted) is now generally available. Google Cloud customers can use the full Claude Platform capabilities with their existing IAM credentials, billing, and committed use discounts, and new features and betas launch on the same day as the native Claude Platform. This version supports Claude Managed Agents (beta), code execution, computer use, Agent skills, MCP connector, Prompt caching, and other capabilities, and provides Claude Console; Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5 are already available, and Claude Opus 5.5's Fast Mode delivers up to 2.5x faster output speed, with input and output pricing at $8 and $40 per million tokens. Anthropic operates this service as the data processor and currently offers US and global endpoints; customers who require Google as the data processor or have regional data residency requirements can still choose the Google-hosted version.

Tomasz Tunguz: Differences in AI Application and Infrastructure Sales—Eight-Figure Deal Closed in Two Weeks

At Fireworks AI, Bardia Shahali grew the sales team from about $3 million to about $400 million; he is now VP of Sales at Granola. He points out that at the AI infrastructure layer, individual deals can reach tens of millions of dollars and close in as little as two weeks, but all inference providers are competing for the same set of customers, making competition fiercer. Because model capabilities reach new heights every 45 days, compensation plans have shifted from annual to quarterly. At the infrastructure layer, FDEs, SEs, and AEs are paired to deliver complex products, while at the application layer, AEs and CSMs deliver software, but with higher productivity and adoption speed.

Code testing startup Blacksmith raises $66 million Series C at a $1 billion valuation

Continuous integration cloud service provider Blacksmith has raised a $66 million Series C round led by Sequoia Capital, with participation from Peak XV Partners, Y Combinator and GV, at a company valuation of $1 billion, bringing total funding to $125 million. The company offers CI cloud services on its own servers; more than 10,000 enterprises currently use it, and weekly code checks have grown from about 5 million in early 2026 to about 90 million, while annual recurring revenue grew 17-fold over the same period. CEO Aditya Jayaprakash said AI coding tools have made writing code easy and the bottleneck has shifted to code verification; this round's funds will mainly be used to buy servers to expand the global CI cloud and to grow the team from about 40 people to 75 by the end of the year.

Anthropic CEO Dario Responds to Safety Concerns in a Long Post, Reveals High-End Model Was Held Back for Four Months

Anthropic CEO Dario posted a long message on X responding to questions from AI safety investigator Nathan Calvin, disclosing for the first time that the company had already trained a high-end model at the Mythos level as early as February this year. Because its cybersecurity defense mechanisms were not yet complete, it was released only to a very small number of partners, and was not opened to the public until the safety guardrails were in place in June—about four months in total. He also listed the costs the company has proactively borne for safety, including a large number of false positives caused by biological risk filtering, pushing for legislation such as California SB 53, and a plan to commit 5% of the company's shares to a public interest fund. Dario said this approach amounts to competing toward a higher standard, while Calvin questioned how much more the company can concede in the race for the sake of safety.

English translation is being prepared.

English translation is being prepared.

DeepSeek open-sources DeepInfer, the inference engine behind its API, with support for both Nvidia and Huawei Ascend

DeepSeek has open-sourced DeepInfer, its in-house inference framework used to serve its models in production. Released under the Apache 2.0 license, it lets developers run DeepSeek-V4.1-Flash on their own infrastructure, with support for subsequent models planned. The framework includes optimized compute kernels and communication libraries such as FlashMLA, DeepGEMM, and DeepEP, supports speculative decoding, asynchronous scheduling, and structured output, and also open-sources the deepseek-recipe library (available in Rust and Python versions) for handling request and response transformations. On the hardware side, it supports Nvidia SM100 (Blackwell) architecture and the Huawei Ascend 950 NPU series from day one; older architectures such as Hopper are not listed. Batch-invariant inference and deterministic sampling are available only on Nvidia hardware, and the published performance data is also based on Nvidia GPUs. The current version supports offline inference but lacks complete production serving components such as KV cache management, cross-machine inference stage separation, and replica load balancing.

Ethan Mollick reviews 30 months of AI development in charts

On the release day of his new book Co-Existence, Ethan Mollick used charts to review the AI progress in the 30 months since his previous book Co-Intelligence was published. He said that since GPT-4, 20 models have successively taken the top spot, and 180 models were stronger than GPT-4 at release, 19 of which can run on modern smartphones; the METR long-task benchmark showed that the amount of human work AI can complete increased from 4 minutes in 2024 to 17 hours in April 2026, and the benchmark was subsequently discontinued because it could no longer measure accurately. He proposed a third mode of collaboration beyond centaur and cyborg: conjurer, that is, summoning agents, assigning tasks, and evaluating results. The charts in this article were completed by Opus 5.5 in Claude Projects orchestrating multiple agents.

English translation is being prepared.

English translation is being prepared.

Gray Swan relaunches Arena red-teaming competition, targeting the AI agent era

AI security company Gray Swan on Tuesday launched a redesigned Arena platform, where contestants compete for prize money by tricking AI models into breaking through their guardrails, adding a sandboxed testing environment for indirect prompt injection. The new platform offers beginner-friendly challenges, training certificates, and a pathway for top performers to enter the Gray Swan red-team community and take part in real-world testing. Gray Swan conducts pre-release stress testing for frontier labs such as Anthropic, OpenAI, and Google DeepMind, and its Arena community has about 15,000 members.

Amazon SageMaker JumpStart 2026 Progress Review

From January to September 2026, Amazon SageMaker JumpStart added more than 100 models from 30 providers, of which Alibaba (Qwen) contributed 29, NVIDIA 17, Mistral 7, Google 6 Gemma 4 variants, IBM 4, and Cohere 4. The new models cover multimodal, speech, image, and video generation, and introduce quantized and compact models such as GLM-5.2 FP8 and JetBrains Mellum2-12B to reduce inference costs.

Artificial Analysis launches HealthBench and HealthBench…

Artificial Analysis has launched leaderboards for OpenAI's two health dialogue benchmarks, HealthBench and HealthBench Professional; the former contains 5,000 conversations with general users and health workers, while the latter consists of 525 conversations all aimed at clinicians. On HealthBench Professional, GPT-6 Astra (max) ranks first with a length-adjusted score of 66.0%, GPT-6.1 Sol (max) is at 65.1%, and Claude Sonnet 5.5 (max with fallback) is at 63.0%. On HealthBench, Claude Sonnet 5.5 (max with fallback) leads at 62.1%, while DeepSeek V4.1 Flash (max) is at 60.9%; Claude Opus 5.5 and Claude Fable 5.1 average about 5,900 characters per answer and lose more than 11 points after length adjustment.

Artificial Analysis Releases AA-PersonalityBench: Measuring AI Models' Communication Styles

Artificial Analysis launches AA-PersonalityBench, using an LLM jury to make pairwise comparisons of answers to the same questions, covering 30 personality traits grouped into 8 indices: Expressiveness, Organization, Informativeness, Transparency, Affinity, Conviction, Imagination, and Slop. The benchmark also programmatically measures 21 communication features, such as response length, structure, punctuation, and emphasis; each trait has 30 questions, for a total of 900 questions, and models answer without a system prompt, with results presented as percentile rankings.