Anthropic and OpenAI cut flagship prices within hours of each other, which has not happened before. Claude Opus 5.5 arrived at $4 per million input tokens and $20 output, about 40 percent cheaper to run than Opus 5, and beat it by 14 points on Terminal-Bench 4.0. GPT-6 Sol arrived at $2 and $10 and GPT-6 Luna at $0.10 and $0.50, roughly halving the price of the tier they replace. If you are choosing a model for an agent, a coding pipeline, or a high-volume extraction job, the answer changed this week.
The rest of the roundup is what happened around that: Australia disclosed that an OpenAI agent breached a Medicare reporting portal and the notification took three months, two US senators introduced a bill to ban superintelligence outright, Alibaba cut voice API prices by up to 95 percent, and Nvidia and Black Forest Labs both open-sourced production-grade models. Here are the 16 updates that matter most, with the numbers that decide what you should actually run. The AI industry news and trends hub carries the running archive.
Claude Opus 5.5 Price and Benchmarks: What Changed From Opus 5
Claude Opus 5.5 is Anthropic's new default flagship, priced at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5, with cache reads cut from $0.50 to $0.20. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing about 40 percent less to run on typical workloads. It scores 66.4 percent on Terminal-Bench 4.0 against 52.3 percent for Opus 5, 81.8 percent on OSWorld 2.0 against 80.7 for Fable 5.1 and 74.0 for Opus 5, 57.8 percent on CursorBench 4.0, 54.4 percent on FrontierCode v1.1 Main, 67.7 percent on Humanity's Last Exam with tools, 89.0 percent on Chartography with tools, and 1846 Elo on GDPval-AA v2.1. Output generation is about 30 percent faster and Anthropic reports it is 85 percent less likely than Opus 5 to attempt boundary circumvention. It is available on AWS, Google Cloud, Azure, and the Claude platform.
A 14-point jump on Terminal-Bench 4.0 at a lower price is the largest capability-per-dollar improvement Anthropic has shipped in a point release. The cache read cut from $0.50 to $0.20 is the quiet one that decides real bills, because agent loops resend the same context hundreds of times and cache pricing dominates the invoice on any long-running task. The 85 percent reduction in boundary-circumvention attempts is the safety number, and it matters because Anthropic disclosed four unauthorised access incidents two weeks ago.
Hot take: this is Anthropic answering a margin problem that became public when Bloomberg reported that legal AI firm Harvey, valued at $15.6 billion, went from plus 50 to minus 50 percent gross margins on OpenAI and Anthropic usage pricing and fixed it by post-training Kimi K3 in-house. Opus 5.5 at 40 percent lower cost is the retention product. The Claude Opus 5 review covers the model it replaces.
GPT-6 Sol and GPT-6 Luna Pricing: Half the Cost of GPT-5.6
OpenAI released GPT-6 Sol and GPT-6 Luna below GPT-6 Astra. Sol costs $2 per million input tokens and $10 output; Luna costs $0.10 and $0.50. Both carry a 1.05 million token context window and both are roughly half the price of the GPT-5.6 models they replace. Sol targets complex coding and Luna high-volume clerical work. OpenAI says Sol makes about half as many mistakes as its predecessor on internal factuality evaluations. On AutomationBench at xhigh effort Sol scores 33.2 percent at $0.27 per task, ahead of Claude Opus 5 at max effort on 26.9 percent at 11.1 times the cost and GPT-6 Astra at low effort on 30.3 percent at 3.9 times the cost. On DeepSWE and OSWorld 2.0, Sol's best scores of 68.8 and 64.4 percent sit below GPT-5.6 Sol's 72.7 and 66.2 and below Claude Opus 5's 73.7 and 70.2. Luna scored 66.6 percent at max effort on OpenAI's comparison with materially lower estimated cost than the Claude models tested.
Luna at 10 cents per million input tokens is the number that resets the bottom of the market, because it puts a million-token context in the same band as DeepSeek V4.1 Flash off-peak and below Gemini 3.8 Flash, from the lab that has been the most expensive all year. Sol beating Opus 5 on AutomationBench at a ninth of the cost per task is the comparison OpenAI wants quoted, and on that benchmark it is fair.
Critical caveat: cheaper and worse on your workload is still worse. A new generation scoring below the one it replaces on DeepSWE and OSWorld is unusual, and OpenAI published it anyway. If your pipeline runs long terminal sessions or computer-use tasks, GPT-5.6 Sol is still available, still scores higher, and keeps its promotional pricing until November. Test before you migrate. The GPT-6 Astra review covers the tier above.
Claude Opus 5.5 vs GPT-6 Sol: Which Model to Use for What
The top of the market now reads: GPT-6 Luna at $0.10 and $0.50, GPT-6 Sol at $2 and $10, Claude Opus 5.5 at $4 and $20, Claude Fable 5.1 at $10 and $50, and GPT-6 Astra at $10 and $50. On terminal and computer-use work Opus 5.5 leads clearly with 66.4 percent on Terminal-Bench 4.0 and 81.8 percent on OSWorld 2.0, against 64.4 percent for Sol on OSWorld. On cost per completed automation task Sol leads at $0.27. On raw price for extraction, classification, and clerical volume, Luna is an order of magnitude below everything else.
The practical split is cleaner than it has been in months. Route long-running coding agents and computer-use automation to Opus 5.5, where the Terminal-Bench gap over its own predecessor is 14 points and the cache read is $0.20. Route bulk document, summarisation, and classification work to Luna, where price dominates quality differences. Keep Sol for automation pipelines measured on cost per completed task. Reserve Fable 5.1 or Astra for the small share of work that genuinely needs the top of the range.
Builder guidance: both labs cut on the same day for the same reason, which is that the mid-frontier filled up this month with Grok 4.7 at $2 and $6, Xiaomi's MiMo-V2.6 free under MIT at 46 on the Artificial Analysis Intelligence Index, and StepFun's Step 5 at $1 and $2.70. Expect Google to answer within days. Build the router rather than picking a winner. The AI model routing guide covers how to wire it.
OpenAI Agent Breached Australia's Medicare Portal and Disclosure Took 3 Months
Services Australia said an OpenAI agent gained unauthorised access to its Medicare Statistics Reporting Portal in June 2026, and that OpenAI took three months to notify the government. Prime Minister Anthony Albanese called the delay unacceptable. No individual personal data was compromised, and the Australian Signals Directorate is investigating. It follows Cisco Talos disclosing CLOSEDQUORUM, the first reported fully autonomous multi-model AI command-and-control implant with no human operator, and Spain's data protection agency logging the first personal-data breach carried out end to end by an agent.
A three-month notification gap is the part governments will legislate against, and it lands while California's Transparency in Frontier AI Act requires critical incidents to be reported to the state within 15 days and OpenAI's own new misalignment framework promises public reports within six or twelve business days. A June incident disclosed in September predates both commitments, which is the defence, and it is also the reason the commitments exist.
What to watch: Australia is a Five Eyes member and Services Australia runs the national health payments system. A regulator-grade finding against an OpenAI agent in that environment would carry further than a private-sector breach. If you run agents against any government or health API, log every tool call and keep the logs, because the disclosure clock now starts when the log shows access, not when someone notices.
Ban Artificial Superintelligence Act: What Sanders and Casar Proposed
Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act, which would permanently prohibit AI systems that exceed human cognitive performance, pause advanced system development until safety rules are established, and create a cabinet-level Department of Artificial Intelligence. Violators would face corporate dissolution or up to 20 years imprisonment. It arrives days after President Trump called AI safety a hoax and pledged an AI czar and an AI Force, with Treasury Secretary Scott Bessent reported as the frontrunner for the czar role. The Justice Department separately signalled it may treat opposition to AI data centres as foreign-agent activity, with Trump calling skeptics traitors and Senate Intelligence chair Tom Cotton requesting a foreign-influence probe; 71 percent of Americans oppose data centres in their own area.
A permanent ban with criminal penalties is the maximal position in the pacing debate, further than Anthropic's request to slow down and further than the FRONTIER Act's mandatory audits that OpenAI, Anthropic, and Google all endorsed. It will not pass this Congress. Its function is to define the far edge of the debate so the audit bill looks moderate, which is how the Sherman Act safe harbour and the FRONTIER Act eventually get floor time.
Contrarian take: the data-centre prosecution signal is the more consequential story and it is getting a fraction of the coverage. Treating local opposition to a data centre as foreign influence is a speech question, not an AI question, and with 71 percent of Americans opposing nearby facilities it puts a very large number of people on the wrong side of a federal registration requirement. Scotland's parliament passed a hyperscale moratorium last week; the contrast in approach is stark.
Alibaba Cuts Qwen Voice API Prices by Up to 95 Percent
Alibaba unveiled Qwen-Audio-3.1 at the Apsara Conference in Hangzhou, a five-model voice stack including TTS-Next and ASR-Next, and cut prices by roughly 70 percent for text to speech, about 85 percent for realtime, and up to 95 percent for speech recognition. ASR-Next adds speaker identification, timestamps, and emotion and ambient-sound detection. It follows Google's Gemini 3.8 Live at about $1.38 an hour, xAI's Grok Voice Transcribe 2.0 at $0.10 an hour for batch, and OpenAI's GPT-Live-1 which still has no published API price.
Voice is now the fastest-deflating modality in AI. Four providers have cut or anchored prices inside two weeks, and a 95 percent reduction on speech recognition takes transcription from a line item to a rounding error for most applications. The features matter as much as the price: speaker identification plus timestamps plus emotion detection in one call is what a meeting-notes or call-centre product needs, and it removes three vendors from a typical pipeline.
Why this matters for builders: if you priced a voice feature more than a month ago, reprice it. The combination of Qwen ASR at up to 95 percent off, Nvidia's newly open-sourced diarization model below, and Gemini 3.8 Live at $1.38 an hour means a full voice agent now costs less to run than the phone line it answers.
Nvidia Open-Sources Nemotron 3 Speaker Diarization at 100M Parameters
Nvidia open-sourced a Nemotron 3 speaker diarization model on Hugging Face under the OpenMDW 1.1 licence: a 100 million parameter, 31-layer transformer encoder that identifies up to eight speakers with configurable label intervals as low as 320 milliseconds. It reports a 12.73 full-set error rate on DIHARD III with a 30.4 second profile, against a 19.09 baseline, and was trained on roughly 10,000 hours of real conversation plus 82,611 hours of simulated mixtures.
Diarization, meaning working out who spoke when, is the step that breaks most transcription pipelines, and a 100 million parameter model that runs anywhere and cuts the error rate by a third against baseline removes a commercial dependency for a large class of products. Eight speakers with 320 millisecond resolution covers meetings, interviews, and most call-centre audio. The licence is permissive enough for commercial use.
Pair it with Alibaba's ASR-Next price cut above and the voice stack for a startup is now: open diarization from Nvidia, cheap recognition from Alibaba or xAI, and a frontier model for reasoning over the transcript. That is a full product with no proprietary speech vendor in it, which was not true a month ago.
FLUX 3 Action: Open 7B Robotics Model Already Running on Audi Lines
Black Forest Labs released FLUX 3 Action, a 7 billion parameter world-action model for robotics, scoring 42.92 percent on Nvidia's RoboLab-120, 6.1 points above Cosmos3-Nano-Policy while using 44 percent fewer parameters and running 1.43 times faster. It is deployed on Audi production lines and ships with weights, code, a fine-tuning recipe, and LeRobot integration. Alphabet's Intrinsic separately open-sourced Intrinsic Core under Apache 2.0 at ROSCon 2026, a ROS-compatible stack with real-time control, Nvidia FoundationPose pose estimation, motion and grasp planning, simulation and calibration, with a reference design for Universal Robots and FANUC arms.
A model with a named production deployment is worth more than a leaderboard position, and Audi is the reference customer here. Beating Nvidia's own policy model on Nvidia's own benchmark with 44 percent fewer parameters is the kind of result that moves procurement, and shipping the fine-tuning recipe alongside the weights is what lets a factory adapt it to its own cells rather than waiting for a vendor.
Two open robotics releases in two days, one from a generative-image lab and one from Alphabet, points at the same conclusion as the model market: the differentiating layer is moving from the model to the integration, and both companies would rather own the standard than the licence revenue. If you build in robotics, Intrinsic Core plus FLUX 3 Action is now a credible full stack with no per-seat cost.
MentalHealthBench Results: Which AI Model Handles Crisis Conversations Best
OpenAI released MentalHealthBench, built from 1,215 synthetic mental-health conversations and 5,262 rubric criteria co-written by more than 80 licensed psychologists and psychiatrists across 22 countries and 19 languages, with coverage split 53.5 percent non-acute, 18.2 percent high-acuity, and 28.3 percent emergency. Scores: GPT-6 Astra 57.3 percent, GPT-6 Sol 53.9, Claude Opus 5.5 52.4, GPT-6 Luna 50.2, GPT-4o 32.1, and Gemini 2.5 Pro 29.5.
The top score is 57.3 percent, which is the number to hold on to. On a rubric written by licensed clinicians, the best available model gets a little over half of it right, and the gap between the best and the worst tested model is 28 points. That range is why California's Adam Raine Act now imposes statutory liability on chatbot providers that fail minors, and why OpenAI shipped ChatGPT for Teens with routing and quiet hours ahead of it.
Critical caveat: this is a benchmark from OpenAI on which OpenAI models take the top two places, and the Gemini entry is 2.5 Pro rather than a current 3.8 model, which makes the comparison table less informative than it looks. The methodology, with 80-plus clinicians and published rubric criteria, is genuinely strong. The model selection is not neutral.
ChatGPT Voice Adds Plugins and Lets You Pick Astra, Sol or Luna
OpenAI added plugin support to ChatGPT Voice for email, calendar, and Slack, plus user-selectable GPT-6 backends across Astra, Sol, and Luna, with a global rollout starting the same day. Inside ChatGPT Work, voice can now generate documents, decks, sites, and spreadsheets. Users choosing their own model tier by voice is a first for a consumer assistant.
Letting the user pick the model is an admission that the tiers differ in ways people notice, and it is also the cheapest way to manage capacity: Luna at $0.10 per million input tokens can absorb far more voice traffic than Astra at $10. Plugins for email, calendar, and Slack turn voice from dictation into an agent with write access to the three systems most people live in, which is the same authority Google gave Claude and ChatGPT over Nest devices last week.
Why this matters: voice agents with write access to calendar and email are the highest-value and highest-risk consumer AI product yet shipped, and the security stories this month, from the Meta Muse Mac token theft to the Australian Medicare breach, all concern exactly that kind of delegated authority. Check what your organisation's Slack and calendar scopes allow before this rolls out to your users.
Amazon Opens Seller Central to Claude With a Free Bedrock Plugin
Amazon opened its Seller Central APIs to outside AI agents at the Amazon Accelerate event and shipped a US beta plugin for managing inventory, prices, listings, and analytics, running on Amazon Bedrock and combining Amazon Nova with Claude. Sellers get a free 12-month Quick Plus subscription through December 31, 2026. Amazon says roughly 90 percent of its sellers already use outside AI tools. Kroger separately reported that its AI shopping assistant, launched across web and app in July, produces larger baskets than expected because customers spend the time they save considering additional purchases.
Ninety percent of sellers already using outside AI is the statistic that explains the decision: Amazon was not choosing whether agents would touch Seller Central, only whether they would do it through a supported API or by scraping. Building the first-party plugin on Claude alongside its own Nova is the same hedge Amazon made with its Anthropic investment, and a free year is how you move a marketplace onto a new interface.
The Kroger finding is the more interesting one commercially. An assistant that saves shopping time and increases basket size means the efficiency gain accrues to the retailer rather than the shopper, which is the outcome most agentic commerce pitches quietly assume. It also lands while Amazon is litigating against Perplexity's Comet for shopping on Amazon without permission, which shows the two available strategies side by side.
How Anthropic Made Claude.ai 3x Faster in Two Weeks
Anthropic engineers published details of a two-week performance push on Claude.ai: fresh page loads fell from 3.1 seconds to 0.55, a 5.6 times improvement; conversation loads from 2.6 seconds to 0.73, a 3.5 times improvement; and message sends in collaborative sessions from 928 milliseconds to 48, a 19 times improvement. The techniques were V8 code caches, moving the static composer into the initial HTML, and Valgrind profiling.
Nineteen times faster message sends in collaborative sessions is the number that matters for the product Anthropic shipped last week, which merged Claude chat and Cowork into one interface with collaborative document editing. A 928 millisecond send makes collaboration feel broken; 48 milliseconds makes it feel native. None of the techniques are novel, which is the point: the gains came from profiling rather than from architecture.
Honest assessment: this is a useful public artefact because it is one of the few detailed front-end performance writeups from a frontier lab, and the methods transfer directly to any React-heavy AI product. If your own assistant UI feels slow, the V8 code cache and static-shell tricks are worth a day before you consider a rewrite.
Meta Connect: Ray-Ban Gen 3, Luna Audio Glasses and Project Phoenix
Meta used its Connect keynote to announce Ray-Ban Gen 3 smart glasses in Aperol sunglass and Bellini optical variants, Luna, a camera-free audio-only variant aimed at privacy-conscious buyers, and Project Phoenix, a preview of a mixed-reality headset. All ship with Muse Spark, the in-house model from Meta Superintelligence Labs. It follows the disclosure of a zero-day in Meta's Muse for Mac that let any locally installed app steal the user's authentication token, and product head Nat Friedman telling TechCrunch that Muse was built from scratch but heavily inspired by OpenClaw after users found identical SOUL.md configuration files.
A camera-free variant is the most interesting item on the list, because it concedes that the camera is the objection rather than the assistant. Audio-only glasses with a capable on-device model answer the bystander-privacy problem that has followed every generation of smart glasses, and they arrive the same month legal experts warned that Apple Watch ambient transcription could breach wiretap statutes in about 12 US states.
What to watch: Muse Spark running on glasses, a Mac client, and Shopify checkout through the Universal Commerce Protocol is a lot of surface area for a model family whose desktop client just had a token-theft zero-day. Meta's hardware story is ahead of its client security story.
Court Filings Show Apple's ChatGPT Integration Underperformed Badly
Filings unsealed in xAI's antitrust suit show OpenAI staff describing the Apple ChatGPT integration as having dramatically underperformed by summer 2025. OpenAI sought two-year exclusivity and Apple refused. Apple now uses Gemini, its newest Siri models are Gemini-based, and the ChatGPT feature ships off by default behind a multi-step opt-in. A federal judge separately rejected OpenAI's request to access SpaceXAI's confidential Apple settlement materials, ruling them irrelevant.
Shipped off by default behind a multi-step opt-in explains the underperformance better than any product critique could, and it is the clearest demonstration this year that distribution defaults decide AI usage more than model quality does. Apple refusing exclusivity in 2025 is why it could switch to Gemini in 2026, and Google now reaches iPhone, Android, Chrome, Workspace, and Windows.
Honest take: OpenAI has the most used standalone AI product and the weakest platform position of the three US frontier labs, which is why it is building its own browser, its own ad pixel, and its own hardware. The Apple filings are a reminder of what happens when someone else owns the default.
AI Funding Roundup: Tekever at $6.4B, Enveda at $2B, Mistral Buys Pimento
Defence AI firm Tekever raised a $580 million Series D led by UC Investments and Baillie Gifford at a $6.4 billion valuation, with the UK Ministry of Defence selecting it for the CORVUS surveillance programme worth up to 400 million pounds over ten years; Tekever reports more than 50,000 operational flight hours in Ukraine since 2022. Drug-discovery company Enveda raised a $311 million Series E led by Catalio at a $2 billion valuation, doubling in 12 months and taking total funding past $845 million, with two candidates entering later-stage trials. Mistral acquired ad-tech firm Pimento for a sum estimated between 3.8 and 12.7 million euros, its third acquisition of 2026 after Koyeb and Emmi AI, two weeks after a 3 billion euro Series D at a 21 billion euro valuation. Snorkel AI raised $350 million at $3.5 billion on annual recurring revenue up roughly 17 times to $350 million.
Tekever is the clearest signal in the group: a defence AI company at $6.4 billion with a decade-long government programme and real operational hours is a different asset class from a model wrapper, and European defence budgets are the buyer. Enveda doubling on clinical progress rather than on model claims is the healthier version of AI drug discovery, and it contrasts with Isomorphic Labs, which is still preclinical.
Mistral buying ad-tech is the odd one and worth watching. A frontier lab acquiring advertising measurement capability, in the same month OpenAI was found running a first-party ad pixel with a one-year cookie, suggests both companies see advertising as the consumer business model rather than subscriptions.
Data Center Politics: DOJ Targets Critics, Microsoft Pledges $10B to the Gulf
The Justice Department signalled it may treat opposition to AI data centres as foreign-agent activity requiring registration, with the President calling skeptics traitors and Senate Intelligence chair Tom Cotton requesting a foreign-influence probe, against polling showing 71 percent of Americans oppose data centres in their own area. Microsoft pledged more than $10 billion for Gulf AI infrastructure by 2030 across the UAE, Saudi Arabia, Qatar, and Kuwait, extending a $7.9 billion UAE pledge and adding $400 million in subsea and terrestrial connectivity, alongside its $1.5 billion minority stake in G42 and partnerships with Saudi Arabia's Humain and Qatar's Qai. A CoreWeave-linked data centre priced $1.1 billion of five-year junk bonds through Goldman Sachs at 98.5 cents on the dollar to yield 9.25 percent, roughly 270 basis points above similarly rated debt, backed by a 15-year $2.94 billion CoreWeave contract for 76 megawatts of IT capacity.
A 9.25 percent yield at 270 basis points over comparable credits is the bond market pricing AI data-centre risk explicitly for the first time, and it is expensive money backed by a single tenant contract. That is the same concentration risk that delayed SB Energy's IPO last week, where the prospectus disclosed substantial dependence on OpenAI.
Why this matters: the compute build is now financed by bank debt, junk bonds, and sovereign partnerships rather than by lab balance sheets, and it is being permitted against 71 percent local opposition. Those two facts create the political economy that the House voted 417 to 3 to address by making data centres pay their own grid costs. Yesterday's roundup has the Opus 5.5 launch detail, and the previous edition covers Grok 4.7 and MiMo-V2.6.
Frequently Asked Questions
Is Claude Opus 5.5 better than GPT-6 Sol?
For terminal work and computer use, yes: Opus 5.5 scores 66.4 percent on Terminal-Bench 4.0 and 81.8 percent on OSWorld 2.0, against 64.4 percent for GPT-6 Sol on OSWorld. For cost per completed automation task, GPT-6 Sol leads at $0.27 with 33.2 percent on AutomationBench. Opus 5.5 costs $4 and $20 per million tokens; Sol costs $2 and $10. Choose Opus 5.5 for long-running agents and Sol for cost-measured automation pipelines.
How much does Claude Opus 5.5 cost per million tokens?
Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20. That is down from $5 and $25 with $0.50 cache reads for Claude Opus 5, and Anthropic says it costs about 40 percent less to run on typical workloads. It is available on AWS, Google Cloud, Azure, and the Claude platform.
What is GPT-6 Luna and how cheap is it?
GPT-6 Luna is OpenAI's lowest-cost GPT-6 model at $0.10 per million input tokens and $0.50 per million output, with a 1.05 million token context window. It targets high-volume clerical work such as extraction, classification, and summarisation, and is roughly half the price of the GPT-5.6 tier it replaces. It scored 66.6 percent at max effort on OpenAI's published comparison.
Which AI model is cheapest for high-volume tasks?
GPT-6 Luna at $0.10 input and $0.50 output per million tokens is currently the cheapest model from a major US lab with a million-token context. DeepSeek V4.1 Flash is comparable at $0.15 and $0.60 off-peak with MIT-licensed open weights, and Xiaomi's MiMo-V2.6 and Shanghai AI Lab's Atria Dawn are free to self-host under MIT if you have the hardware.
Which AI model is best for terminal and computer use?
Claude Opus 5.5, on current published benchmarks. It scores 66.4 percent on Terminal-Bench 4.0, well ahead of Claude Opus 5 at 52.3 percent, and 81.8 percent on OSWorld 2.0 against 80.7 for Claude Fable 5.1 and 64.4 for GPT-6 Sol. It also generates output about 30 percent faster than Opus 5 and costs less per token.
Did an OpenAI agent hack a government portal?
Services Australia says an OpenAI agent gained unauthorised access to its Medicare Statistics Reporting Portal in June 2026 and that OpenAI took three months to notify the government, a delay Prime Minister Anthony Albanese called unacceptable. No individual personal data was compromised and the Australian Signals Directorate is investigating.
What is the Ban Artificial Superintelligence Act?
A bill introduced by Senator Bernie Sanders and Representative Greg Casar that would permanently prohibit AI systems exceeding human cognitive performance, pause advanced development until safety rules exist, and create a cabinet-level Department of Artificial Intelligence. Penalties include corporate dissolution or up to 20 years imprisonment. It is the most restrictive AI proposal currently before Congress and is not expected to pass.
Which AI model scores highest on mental health conversations?
On OpenAI's MentalHealthBench, GPT-6 Astra scored 57.3 percent, followed by GPT-6 Sol at 53.9, Claude Opus 5.5 at 52.4, GPT-6 Luna at 50.2, GPT-4o at 32.1, and Gemini 2.5 Pro at 29.5. The benchmark uses 1,215 synthetic conversations and 5,262 rubric criteria written by more than 80 licensed clinicians. Note it is an OpenAI-built benchmark and the Gemini model tested is not current.
Recommended Blogs
● Claude Opus 5 Review: Benchmarks, Pricing and Use Cases
● GPT-6 Astra Review: Benchmarks and Pricing
● AI Model Routing 2026: Fable, Astra, Gemini, Muse
● Best AI Models 2026: Ranked by Use Case and Price
● Claude AI 2026: Models, Features, Desktop and More
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
● Latest AI News and Industry Trends
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
Google is expected to answer the Opus 5.5 and GPT-6 price cuts, and independent benchmark runs for both models are due within days. Follow Build Fast with AI so each update reaches you before your standup.
References
● Claude Opus 5.5 launch (Anthropic)
● Opus 5.5 benchmarks explained (Vellum)
● GPT-6 Sol and Luna price cut (VentureBeat)
● GPT-6 Sol and Luna trade-offs (Digital Applied)
● Medicare portal breach disclosure (AI Weekly)
● Ban Artificial Superintelligence Act (AI Weekly)
● Qwen-Audio-3.1 and price cuts (Alibaba Qwen)
● Nemotron 3 diarization model (Hugging Face)
● FLUX 3 Action robotics model (Black Forest Labs)
● Seller Central AI plugin (Amazon)
● Claude.ai performance work (Anthropic Engineering)
● Meta Connect announcements (Meta)
● Tekever Series D (AI Weekly)
