Max 20x becomes Ultra with no 5-hour limit for verified accounts, Cursor pulls backdoored Claude Code extension, DiffuSpace releases Clover-35B-A3B

Anthropic renames its Max plans: Max 5x becomes Max, Max 20x becomes Ultra, at the same prices, and the 5-hour limit disappears for verified Ultra accounts. Cursor removes backdoored versions of the Claude Code extension from its extension marketplace, which may have affected at most 5,000 users, and DiffuSpace releases Clover-35B-A3B-Preview, an open-weights diffusion language model. OpenAI also launches GPT-6 Rosalind Research for its approved life sciences customers, while Claude Code 2.1.314, Antigravity’s plugin Marketplace, and Kimi’s 3D Make round out the day.

Max 5x becomes Max, Max 20x becomes Ultra, and the 5-hour limit disappears for verified Ultra accounts

October 30 — Anthropic renames its two Max plans. Max 5x is now simply called Max, and Max 20x becomes Ultra. Prices remain unchanged at 100 dollars per month for Max and 200 dollars for Ultra, and existing subscribers are transitioned automatically, with no action required.

We’re renaming our Max plans: Max 5x → Max, Max 20x → Ultra. Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do. — @claudeai on X

The name change comes with a new feature for the most expensive plan: the 5-hour limit is removed for verified Ultra accounts, “starting today for most existing subscribers.” Newer accounts retain a temporary limit until verified—which Anthropic justifies by fraud and abuse prevention—and receive an email when it is lifted. Boris Cherny, the creator of Claude Code, reshared the announcement a few minutes later.

That leaves the weekly cap, on which sources say two things that should be read side by side. The tweet claims that “weekly limits remain unchanged.” The Ultra plan help page, updated the same day, states for its part that Ultra includes 10 times the usage of Pro each week, and the pricing page echoes this “10x Pro usage every week,” for a plan that was until now called Max 20x. Users asked about this under the tweet; Anthropic had not replied at the time of our report.

Plan (new name)	Former name	Monthly price (web)	5-hour limit	Displayed weekly usage	Monthly API credits
Max	Max 5x	100 dollars (€90 on the French page)	Yes (5x Pro every 5 hours)	5x Pro	100 dollars
Ultra	Max 20x	200 dollars (€180 on the French page)	No, for verified accounts	10x Pro	200 dollars

The monthly API credits announced in early October follow the new names. According to the help page, the 200-dollar price is for the web, on a monthly subscription only, with mobile pricing varying depending on the platform.

🔗 The Ultra plan help page · Claude’s pricing page · Boris Cherny’s post

Cursor removes backdoored versions of the Claude Code extension from its marketplace, at most 5,000 users affected

October 29 and 30 — Cursor reported a security incident on its own VS Code extension marketplace. On October 29, between approximately 2:27 AM and 1:35 PM (Pacific Time), the marketplace offered modified, malicious versions of the Anthropic.claude-code extension for Windows and macOS under numbers 2.1.312 and 2.1.313. According to Cursor, they were not developed by Anthropic. 2.1.312 reuses the number of a genuine Claude Code release, published by Anthropic on October 27; Anthropic has never published a 2.1.313 version, either on npm or on GitHub.

We identified a security incident affecting a modified Claude Code VS Code extension accessed through our Cursor IDE extension marketplace, impacting at most 5,000 users. This malicious version of the extension was not developed by Anthropic. — @cursor_ai on X

The backdoored extension attempted to install a persistent SSH backdoor. It sent machine information to the attacker (hostname, user, operating system, environment variable names, working directories, Git email, public IP), added its SSH key to authorized keys, and then attempted to enable the SSH service and open a reverse SSH tunnel, relaunched approximately every 60 seconds by a LaunchAgent on macOS or a scheduled task on Windows. On most machines, opening the tunnel failed due to a lack of administrator privileges and errors in the malicious code: the attacker was therefore unable to connect to them. The key, persistence, and sending of information, however, did take place.

Cursor says it removed the version as soon as it was detected, shut down the entire marketplace as a precaution, verified that no other extensions had been modified, rotated its keys, and then rebuilt the marketplace before reopening it. The attacker’s server was taken offline in coordination with its host. Cursor says it has no evidence that other versions of the extension or the official distribution are affected. Its advisory, updated on October 30 to clarify indicators of compromise and affected versions, summarizes the incident:

Incident element	What Cursor says
Backdoored versions	Anthropic.claude-code 2.1.312 and 2.1.313 (Windows, macOS)
Exposure window	October 29, approximately 2:27 AM to 1:35 PM PT
Users affected	At most 5,000 (tweet)
Indicator of compromise	“svc-telemetry” SSH key in authorized_keys
Remediation tools	remediate-macos.sh and remediate-windows.ps1 scripts
Version to install	Claude Code for VS Code 2.1.314 (tweet)

According to Cursor, uninstalling the extension may not be enough: those who installed or updated it during this window must also rotate credentials and sessions accessible from the machine. Cursor is asking all its customers to upgrade to 2.1.314, detailed below in the coding agents section. Anthropic had not commented on the incident at the time of our report.

🔗 Cursor’s security advisory

DiffuSpace releases Clover-35B-A3B-Preview, an open-weights diffusion language model

October 30 — DiffuSpace, where members of the Dream 7B team (released in April 2025 by the University of Hong Kong’s NLP group) are continuing their work, presents Clover-35B-A3B-Preview, the first release in its Clover family. It is a diffusion language model: instead of writing one token after another, it denoises entire blocks in parallel. It is a 35-billion-parameter MoE with 3 billion active parameters, featuring a thinking mode (thinking mode), released under the Apache 2.0 license on Hugging Face and ModelScope, with code on GitHub and support for SGLang and llama.cpp. According to its Hugging Face model card, it is derived from Qwen3.6-35B-A3B.

On agent tasks, DiffuSpace reports the highest scores among the models it compares across six benchmarks, ahead of two other diffusion models (LLaDA2.2 flash, Mercury 2.5) and ahead of the autoregressive models Qwen3.6-35B-A3B and Qwen3.5-35B-A3B. The figures below are from DiffuSpace:

Benchmark cited by DiffuSpace	Clover-35B-A3B-Preview score
SWE-bench Verified	73.6
Terminal-Bench 2.1	51.7
τ³ Overall	65.3
BFCL-v4	67.9
GPQA Diamond	83.8
MMLU-Pro	84.8
ZebraLogic	95.8
AA-LCR	72.0

The advantage of diffusion lies in speed, again according to its authors:

Measurement setup	Clover throughput	Comparison cited by DiffuSpace
One GB300, BF16, batch size 1 (internal agent workloads)	680 tokens/s	—
H200 (agents)	515 tokens/s	365 tokens/s for a similarly sized Qwen with DFlash2
16 GB RTX 5070 Ti, INT4 (MATH-500)	101.8 tokens/s	48.5 tokens/s for similar MiniCPM and Qwen models

To achieve this, the team redesigned the Gated DeltaNet architecture for block diffusion: the previously generated context is compressed into a recurrent state, while tokens in the current block are read bidirectionally. DiffuSpace also announces, for later release, a decision model derived from Clover that answers up to eight questions in a single denoising step, as well as an agent harness and a desktop coding client, both still in development.

🔗 DiffuSpace’s blog post · The model card

OpenAI launches GPT-6 Rosalind Research, restricted to approved life sciences customers

October 30 — OpenAI deploys GPT-6 Rosalind, the next generation of its model specialized in life sciences, named GPT-6 Rosalind Research in ChatGPT. The API changelog announces the identifier gpt-6-rosalind, which combines specialized biological reasoning with scientific tool use. Access remains behind the trusted-access program (trusted-access program): only organizations approved for Rosalind Research can use it, for internal research conducted by authorized users.

In ChatGPT, the rollout targets a limited set of approved Enterprise customers, in Work and Codex, but not in the Chat tab. It begins with a two-week admin preview, disabled by default, and existing Research approval does not exempt organizations from this step. On the API side, usage is limited to internal research tools, with no availability on Amazon Bedrock. A Discovery tier, for routine scientific work, remains on the Rosalind 5.5 model.

Billed model	Input price (per million tokens)	Output price	Cached input
`gpt-6-rosalind`	15 dollars	75 dollars	1.50 dollars
`gpt-rosalind-research`	5 dollars	25 dollars	0.50 dollars

The new model costs three times the price of gpt-rosalind-research. OpenAI had published neither a blog post nor benchmarks at the time of our report.

🔗 ChatGPT Enterprise and Edu release notes · The API changelog · The API pricing page · The GPT-Rosalind help page

Coding agents: Claude Code 2.1.314, Antigravity 2.28.2 plugin Marketplace, and Antigravity CLI 1.4.10

Claude Code 2.1.314 blocks writes outside working directories and extends idle compaction to VS Code and Desktop

October 30 — Claude Code 2.1.314 is a major maintenance release: 119 entries, including 81 fixes. It is also the version recommended by Cursor following its marketplace incident. The new setting permissions.blockWritesOutsideWorkingDirectories prevents the Write, Edit, and NotebookEdit tools from writing outside working directories, across all permission modes.

What’s new in 2.1.314	What it changes
`permissions.blockWritesOutsideWorkingDirectories`	Write, Edit, and NotebookEdit reject paths outside working directories
Idle compaction	Extended to VS Code and the Claude Desktop Code tab, beyond 200K tokens
`groups` (Claude apps gateway)	Upstream restricted to listed IdP groups
`guardrail_trace` (Bedrock upstreams)	Logs the triggered guardrail filter, without the detected text
Admin-capped effort	Effort automatically lowered and reported, instead of the “Invalid effort level” error

Among the fixes, claude -p now exits with code 1 and displays the error when startup fails, instead of 0 silently, and auto-resume upon hitting usage limits no longer retries more than three times without intervention. Anthropic released no 2.1.313 version: numbering jumps from 2.1.312 to 2.1.314. 2.1.315 has been on npm since the evening of October 30, with no release notes at the time of our report.

🔗 Claude Code 2.1.314 release notes

Antigravity 2.28.2 opens the plugin Marketplace to third-party partners

October 29 — Google opens Antigravity’s plugin Marketplace to third-party publishers, both in the application and in Antigravity CLI. The blog post calls it a launch, but a plugin marketplace had already existed since version 2.18.1 on September 28: the novelty lies in the partners. Twelve are showcased, including Atlassian, CrowdStrike, Databricks, Datadog, Figma, Stripe, and Unity. A plugin is a versioned package, described by a plugin.json manifest, bundling up to five building blocks installed and activated together: skills, MCP servers, rules, hooks, and custom agents.

Application 2.28.2 organizes these plugins into shelves by category and adds a Featured section. An Accounts tab manages connections for plugins that require one, and the agent can suggest in conversation a plugin found in the Marketplace; these two features are not yet available for business and enterprise accounts. In the CLI, installation is done via agy plugin install.

🔗 Antigravity’s blog post

Antigravity CLI 1.4.10 Lightens Continuous Display by 70%

October 29 — Antigravity CLI updates to 1.4.10. Lines that move off-screen now join the terminal’s scrollback history instead of triggering a full redraw, reducing written data volume by about 70% and making the CLI smoother over a slow SSH connection. About 100 MB of memory is returned to the system 15 seconds after opening, and attached images are limited to 2,000 pixels per side to prevent overly heavy requests. Sub-agents no longer lose their assigned task after compaction.

The release caps a series that our tracker had not covered:

CLI Version	Release Date	Key Change
1.4.10	October 29	Lighter scrolling, memory returned at startup, resized images
1.4.9	October 28	Native image pasting, even over SSH; separate authorization for Gmail, Docs, Drive, and Google Cloud Console
1.4.8	October 27	Turn-by-turn navigation in /diff, longer Auto mode command review

🔗 Antigravity changelog

OpenAI Promises a Semiannual Community Safety Report and Tightens Protections for Minors

October 30 — OpenAI announces two measures for ChatGPT. The first is a transparency commitment: a Community Safety Report, dedicated to violence prevention and the effectiveness of guardrails. The first edition will be published in early 2027 and will cover the second half of 2026, after which the report will be updated every six months. OpenAI indicates that it referred more than 200 reports to law enforcement in the third quarter of 2026 for imminent threats of violence.

The second tightens the rules. For all users, ChatGPT will provide fewer details on past mass casualty attacks, without reconstructing them (weapons, equipment, how to obtain them) or comparing their scale, and will be able to end a conversation itself if marked by a fixation on violence; multiple terminations of this type can lead to a temporary account suspension. For users under 18, restrictions extend to actionable information about weapons and security systems, even without apparent violent intent, and age prediction continuously reassesses new signals. These protections will roll out in the coming weeks.

🔗 OpenAI blog post

FrontierCode 2.0: Gemini 4 Carbon Ranks Fourth, at 40% for 4.04 Dollars per Task

October 30 — Cognition added Gemini 4 Carbon to FrontierCode 2.0, its benchmark evaluating whether a PR would actually be merged by the repository maintainer. Google DeepMind’s model takes fourth place with 40.0%, at medium effort in the Devin CLI harness. The reported cost is the average expenditure per run at the public API rate.

Evaluated model (effort)	FrontierCode 2.0 score	Cost per task
Claude Opus 5.5 (max)	45,5 %	12,82 dollars
GPT-6 Astra (max)	42,8 %	6,68 dollars
Claude Sonnet 5.5 (max)	41,0 %	19,94 dollars
Gemini 4 Carbon (medium)	40,0 %	4,04 dollars
GPT-6.1 Sol (max)	39,0 %	1,03 dollar
Gemini 4 Argon (high)	26,3 %	6,51 dollars

Gemini 4 Argon remains twelfth, behind Claude Haiku 5.5 and Grok 4.7 but ahead of Muse Wonder 1.0. Its cost differs depending on the source: 6.51 dollars per task on the leaderboard page, 6.41 dollars in Cognition’s tweet.

🔗 Cognition’s tweet · FrontierCode leaderboard

Creative Agents: Runway Agent Opens Files and Executes Code, Kimi Launches 3D Make

Runway Agent Opens Files and Executes Code

October 30 — Runway expands the role of Runway Agent, its conversational creative agent. Previously focused on generation, it can now, according to Runway, open files and execute code.

Now, Agent does more than generate. It can open files and run code, it builds your brands launch kit and creates product shots, ad variations, motion graphics from a reference and more. — @runwayml on X

The announcement consists of a tweet and a demo video: no blog post, no changelog entry, and no details regarding eligible plans, the code execution environment, or credit costs. The agent is accessible from the Agent page of the Runway application.

🔗 Runway’s link to the agent

Kimi Launches 3D Make, Which Creates 3D Models, Scenes, and Games

October 30 — Moonshot AI officially unveils 3D Make, Kimi’s 3D creation space, available on kimi.ai and in Kimi Work. Mentioned in passing alongside Kimi K3.1 Pro on October 23, the tool now gets a help page and a tutorial. Powered by K3.1 Pro, it offers three types of creation:

Creation Type	What 3D Make Produces
3D Web	Interactive 3D web sites, scenes, and games, with preview and sharing
3D Modeling	Models from text, a sketch, or a reference image
3D Printing	Files ready for 3D printing

According to Moonshot, K3.1 Pro was specifically trained to control 3D software via Python scripts, which also opens up tools without an MCP server; it hands off characters or statues to meshing services like Tripo, Meshy, or Hyper3D. Kimi is adapted for Unreal Engine, Unity, and Godot, but plugins requiring locally installed software only work in Kimi Work. 3D Make draws from Kimi’s unified usage quota, with no separate pricing. Moonshot acknowledges its limitations: character animation (rigging), complex materials, physics simulation, and the consistency of entire 3D worlds.

🔗 Moonshot AI’s tweet · 3D Make help page · Kimi Academy tutorial

Perplexity Connects Computer’s Mac and Windows Apps to Chrome

October 30 — Perplexity releases a new Chrome extension connecting the Mac and Windows apps of its agent, Computer, to the user’s browser. Computer can thus work on websites where the user is already logged in. In the demo, it turns reviews left for a business into a prioritized list of fixes and replies. The extension is available now for all Computer users in the Mac and Windows apps, via the Chrome Web Store.

This is not a first: since October 5, Mac app 26.38.0 could already use Chrome, Edge, or Brave via an extension, with Comet remaining the default browser. What’s new is the Windows app and the rollout to all Computer users. Perplexity has published neither a version number nor documentation.

🔗 Perplexity’s tweet

NVIDIA Releases RE-USE, an Open 9.6-Million-Parameter Model That Cleans Up Speech

October 30 — NVIDIA highlights NemotronLabs RE-USE, an open, multilingual speech enhancement model designed by its researchers. With 9.6 million parameters, it removes background noise, echo, and other distortions (reverberation, clipping, codec artifacts, packet loss, low-quality microphones). It accepts sampling rates from 8 to 48 kHz and is released under the OpenMDW-1.1 license, ready for commercial use according to its model card. The card has been on Hugging Face since July 22 and was modified on October 22; NVIDIA presents the model as new on October 30.

On the non-blind test set of URGENT 2025, the model card compares it to much larger models, and it does not lead on all metrics:

Evaluated model	Number of parameters	DNSMOS	NISQA	UTMOS	PESQ	STOI	SpeechBERTScore
Sidon	250M	3,34	3,87	3,21	1,73	0,62	0,85
UniPASE	500M	3,21	4,03	2,96	2,17	0,76	0,89
NemotronLabs RE-USE	9,6M	3,29	4,33	2,94	2,57	0,78	0,88

RE-USE achieves the best scores in NISQA, PESQ, STOI, speaker similarity, and transcription accuracy, but remains behind Sidon in DNSMOS, behind Sidon and UniPASE in UTMOS, and behind UniPASE in SpeechBERTScore.

🔗 NVIDIA’s tweet · The model card

Together AI Opens Inference Endpoints Processed Entirely in the United States

October 29 — Together AI launches inference endpoints tied to a geographic region, starting with the United States. By replacing the API base address with us-api.together.ai/v1, using the same key and model identifiers, an enterprise gets the guarantee that its requests are processed 100% in the United States. The service costs 1.25 times the standard serverless rate, for input, cached input, and output tokens.

Seven open models are supported at launch: Kimi-K3, GLM-5.3, GLM-5.3-Flash, DeepSeek-V4.1-Flash, DeepSeek-V4.1-Flash-1017, DeepSeek-V4-Pro-0813, and MiniMax-M3.1-Flash. The documentation specifies the limitations: the guarantee covers the location of request processing, not data storage, log writing, or network routing. If a model is unavailable in the region, the request fails (error region_unavailable) instead of being routed elsewhere.

🔗 Together AI’s tweet · Documentation

Cohere Increases Its Commitment in Canada to Over 4 Billion Canadian Dollars and Plans More Than 1,500 Jobs by 2031

October 30 — Following the initial close of its Series E, Cohere details its commitments in Canada. The company plans to invest more than 4 billion Canadian dollars in R&D and AI infrastructure in the country between 2024 and 2031, across several provinces: this is an update to a multi-year effort that began in 2024, supplemented by new commitments.

Cohere’s Commitment in Canada	Announced Value
R&D and infrastructure investment	More than 4 billion Canadian dollars, 2024-2031
Jobs created	More than 1,500 by 2031, including about 1,000 technical roles
Paid student positions	More than 250 by 2029
GPU partnership with Bell AI Fabric	309 million Canadian dollars, already underway

The academic partnership network builds on recent collaborations with the University of Toronto and the University of Waterloo, and draws on Amii, Vector, and Mila. The partnership with Bell AI Fabric is intended to run Cohere’s models on sovereign Canadian infrastructure. The post does not break down the 4 billion between R&D and infrastructure and makes no numerical connection to the Series E.

🔗 Cohere’s blog post · Cohere’s thread on X

In Brief

APIs and Open Models

Images in the Claude API — Beyond 20 images in a single request, each image can now measure up to 3,000 pixels, compared to 2,000 previously; the 128 MB maximum size applies to Anthropic’s Claude API, with the limit remaining 20 to 32 MB via cloud platforms. 🔗 source
Le Chonk — Mistral shares Thibault Jaigu’s post, which ranks the Mistral Large 4 update second among open models on an unnamed coding benchmark, and thanks Requesty for its testing; no scores were published. 🔗 source
Lumma-Fev-v2 — FrontiersMind releases its 4B and 27B decision models under Apache 2.0: an 87.3% average across eight public benchmarks for the 27B and 82.4% for the 4B, compared to 81.2% for Jev 1.13.0, which retains the edge on BBH and JudgeBench, according to the authors. 🔗 source
Qwen3.8-27B on a 16 GB Mac — Project Pyrite downsizes the model from 51.75 to 10.11 GiB and runs it at 12.2 tokens/s on an M2 Pro with mlx-lm; allocating 3 bits to the 30 most sensitive projections reduces KL divergence by 22% at an equivalent size, according to the author. 🔗 source
NVIDIA’s MonsterHouse — A talking haunted house for Halloween runs on a single workstation with Nemotron Voice Agent, Nemotron 3.5 Bolt E4B, Magpie TTS, and Audio2Face-3D: 1.70 s median latency, 0.72 s for cached responses. 🔗 source

Agents and Coding Tools

Claude Code Consumption — Boris Cherny concludes the investigation: a few minor issues are being fixed, but neither a single cause nor a regression; the accelerated usage mostly stemmed from one or two heavy sessions (effort xhigh, /goal, Workflows, Agent Teams). The next day, he announced CLAUDE_CODE_SUBAGENT_EFFORT to set sub-agent effort. 🔗 source
Vibe CLI 3.0.4 — Mistral adds /release-notes and two output caps, agent_max_output_tokens and compaction_max_output_tokens; without a “breaking” label, custom Python tools from ~/.vibe/tools/ are no longer loaded (migration to a skill proposed), and three network timeouts no longer have any effect. 🔗 source
Kimi Code 2.4.1 — A file read by the agent returns at most 100,000 characters or about 25,000 tokens, down from 500,000 characters previously, configurable via max_chars; tool and sub-agent prompts have been shortened. 🔗 source
ZCode v3.17.1 — First GitHub release since v3.14.3, featuring macOS (arm64, x64) and Windows (arm64, x64) installers and a SHA256SUMS file; no new features described. 🔗 source
Replit — GPT-6.1 Terra joins the Agent’s model picker in Power mode, and the Agent can send a message to another conversation or create one (Pro and Enterprise); no pricing mentioned. 🔗 source
Gemini CLI, October 30 Nightly — The gemini extensions validate command checks an extension before installation or publication, and 429 errors are displayed on the very first attempt with a countdown; stable v0.66.0 remains unchanged. 🔗 source
Antigravity Python SDK Bumps to 0.1.23 with Programmatic Tool Calls in Workflows — Catch-up from October 26: beta.workflows.call_tool, multi-turn workflow agents, structured arguments, and conversation caching for local LiteRT models. 🔗 source
DeepSeek Harness 0.2.1-rc.5 — A new release candidate whose sole change improves reasoning display in conversations; the 0.2.1 series still lacks a stable release. 🔗 source

Image, Video, and Voice

TELUS Digital and ElevenAgents — During a proof of concept of around 2,300 calls, the voice agent calling new internet customers kept 30-day cancellations under half the rate of new customers overall, preserving 55,000 Canadian dollars in revenue per 1,000 contacted customers. 🔗 source
SFX v3 Post — ElevenLabs states that its sound effects model outperforms SFX v2 in benchmark tests, without figures, with fewer artifacts on gongs, chimes, and whistles; it is available in ElevenCreative, the mobile app, and ElevenAPI. 🔗 source
Sales Agents — Catch-up from October 27: according to three studies by ElevenLabs, nearly half of buyers choose voice after trying it; 56% entrust a delivery delay to the agent, 34% a billing error, and 60% would abandon it without visible access to a human. 🔗 source
LoRA for FLUX 3 Image Klein — Black Forest Labs publishes a guide to training a first LoRA (GPU with at least 24 GB of VRAM, ostris AI Toolkit), illustrated with Halloween postcards. 🔗 source
Kling AI Originals — Kling launches a series of short films made with its tools; the first, “My Favorite Night of the Year” by Alex Duloz, was created using Kling 4.0. 🔗 source

Business, Law, and Communities

Credit Saison — Catch-up from October 29: the 315-participant ChatGPT Enterprise pilot shows a 954% return on investment and 14 hours saved per month per participant, according to the company, which subsequently deployed the tool to all its employees. 🔗 source
robots.txt Signal — OpenAI will honor a shared signal in robots.txt, originating from the IETF AI Preferences group, which expresses website preferences regarding foundational generative AI training; implementation is targeted in the coming months. 🔗 source
Fair Use — In the lawsuit brought against it by the New York Times, OpenAI published excerpts from 12 briefs supporting fair use, including those from the United States government and the Authors Alliance. 🔗 source
Scav — OpenAI Developers showcases the Scav robotics community, where GPT-6 Astra remotely controlled a robotic arm; no metrics or products mentioned. 🔗 source
Fine-Tuning at Perplexity — A guide advises refining the prompt, then adding RAG, before moving to fine-tuning, noting that with QLoRA a 7B model can be fine-tuned on a consumer 16 to 24 GB card; no new product announcements. 🔗 source

What It Means

Anthropic is turning the 5-hour limit into a selling point for its most expensive plan: a verified Ultra subscriber is no longer cut off after 5 hours and manages their usage across the week. The key metric to watch is precisely this weekly cap, which the tweet says is unchanged and which the help and pricing pages list at 10 times the usage of Pro. Consumption was already the topic of the week: Boris Cherny attributes the accelerated usage to a few heavy sessions rather than a regression, and adds a setting to reduce sub-agent effort. Uninterrupted time is expanding; the weekly budget is becoming the metric that matters.

The Cursor incident shows that extension marketplaces have become an attack surface for coding agents: an extension bearing the name of a well-known publisher, with a plausible version number, managed to affect up to 5,000 users in about eleven hours. On the same day, Antigravity opened its Marketplace to third-party publishers, with plugins embedding MCP servers, hooks, and agents, while Claude Code added a setting prohibiting any writing outside working directories and Antigravity CLI separated permissions for sensitive Google apps. The more these tools open up to extensions, the more verifying their provenance matters.

Together AI and Cohere are responding to the same demand: knowing where models run. Together charges 25% more to guarantee processing in the United States, without fallback outside the region, on open models that all originate from Chinese labs; Cohere values its commitment to Canada at over 4 billion Canadian dollars, partnering with Bell AI Fabric for sovereign infrastructure. GPT-6 Rosalind Research illustrates another form of control, through access: a model restricted to approved organizations, priced at $15 per million input tokens, with no public blog post or benchmarks.

On the performance side, efficiency is taking precedence over size. Clover relies on diffusion to reach 680 tokens/s on a single GB300, claiming agent scores higher than the Qwen model it derives from; RE-USE cleans up speech with 9.6 million parameters compared to 250- and 500-million models; Project Pyrite fits a 27-billion-parameter model onto a 16 GB Mac. These figures come from the authors, with no independent benchmark to date. FrontierCode 2.0 shows the same trade-off between score and cost: Gemini 4 Carbon reaches 40% for $4.04 per task, compared to 45.5% for $12.82 with Claude Opus 5.5, and GPT-6.1 Sol reaches 39% for $1.03.