OpenAI cuts GPT-6 prices in half

Good day, AI enthusiasts!

It was another busy week in AI, but a few themes stood out.

The money around AI was hard to ignore — but some of the most interesting news had little to do with fundraising.

Here are 10 AI developments worth knowing:

Anthropic says roughly 950 Claude agents searched more than 200,000 reverse transcriptases and identified an unusual combination of a reverse transcriptase, an accessory protein and repeated DNA sequences. Human researchers then investigated the candidate in the lab and named the system array-associated reverse transcriptases, or ART.

The arrangement has features reminiscent of CRISPR and other programmable biological systems, but Anthropic is being careful about what it claims: researchers do not yet know what ART actually does. That distinction matters. The interesting part is not that Claude has suddenly “discovered the next CRISPR”, but that an AI system helped narrow a very large biological search space to something scientists considered worth testing.

Source: Anthropic

Akamai announced a seven-year, $11.6 billion contractual commitment from Anthropic to use its distributed cloud infrastructure for growing CPU workloads. The agreement also allows for up to another $9 billion of expansion, taking the potential relationship to roughly $20 billion.

Most discussion of frontier AI infrastructure understandably centres on GPUs. This deal is a reminder that running a large AI company also creates enormous demand for the surrounding compute, networking and software infrastructure. Akamai getting a contract of this size also broadens the list of companies benefiting directly from frontier-lab spending.

Source: Akamai announcement

SoftBank sold $11.1 billion of dollar- and euro-denominated debt, with Reuters reporting that the money will help fund its OpenAI stake and other AI investments. The bonds carry yields ranging from 7.125% to 9.75%.

The striking part is where AI financing is heading. Huge infrastructure commitments are increasingly being supported not just by cash flow and equity, but by large-scale borrowing. Reuters says SoftBank has now committed $64.6 billion to OpenAI and will own roughly 13% of the company.

Source: Reuters

Oracle issued a force majeure notice relating to Project Jupiter, a 1,400-acre New Mexico data-centre campus connected to the broader Stargate infrastructure programme. Reuters reports that the issue involves potential delays in securing power, although Oracle says the project remains on schedule.

Oracle is not walking away from the project. The notice appears designed to protect it financially if power delays push back the planned 2028 opening. But the episode neatly illustrates one of the less glamorous constraints on the AI build-out: raising billions is one challenge; securing enough electricity, permits, water and physical infrastructure is another.

Source: Reuters

OpenAI expanded the GPT-6 family with Sol and Luna, bringing some of the improvements introduced with GPT-6 Astra to cheaper tiers. API pricing is 50% lower than the promotional pricing of their GPT-5.6 predecessors: Sol drops to $2 per million input tokens and $10 per million output tokens, while Luna falls to $0.10 and $0.50.

OpenAI says both models also improve factual reliability, coding and computer use. The broader commercial point is straightforward: the frontier is increasingly competing on useful intelligence per dollar, not simply who can publish the highest benchmark score.

Source: OpenAI

Anthropic released Claude Opus 5.5 on 22 September, saying it performs around the level of Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads. Token pricing falls to $4 per million input tokens and $20 per million output tokens, while Anthropic says output generation is more than 30% faster.

There is also an interesting safety angle. Anthropic says Opus 5.5 recorded its strongest result yet on the company's automated behavioural audit and improved resistance to prompt injection. Benchmark claims still come largely from Anthropic's own testing, but the pricing move reinforces the same pressure visible at OpenAI: capable models are becoming materially cheaper to deploy.

Source: Anthropic

Meta used Connect 2026 to push its Muse agent further into hardware. The company is working on bringing Muse to its smart glasses, where users will be able to speak to the agent while it handles tasks in the background.

Meta also showed Muse Charm, a small standalone device that does not require users to unlock a phone or open an app before speaking to Muse. Only a handful of prototypes currently exist, according to Mark Zuckerberg, but Meta says it wants to ship the device in December. This is one of the more aggressive attempts yet to turn a general AI agent into something people carry — or wear — throughout the day.

Source: Meta Connect 2026 announcements

Alibaba introduced the Zhenwu V900, which it describes as China's most powerful AI chip. The company says it delivers three times the performance of the previous M890 and can be deployed in clusters containing up to 500,000 cards.

Those are Alibaba's own performance claims rather than independently verified comparisons, and the company has not published an absolute FLOPS figure. But the accompanying roadmap is significant: Alibaba is training Qwen 4, planning future models reaching 5–10 trillion parameters, and targeting more than 20GW of global data-centre capacity by 2032.

Source: Alibaba Group

Politico reported that the White House asked OpenAI and Anthropic to withhold new models from British testers until the systems had first undergone a US review. Reuters, via CNA, reported that the request is tied to concerns about the security of increasingly capable AI systems.

At the time of the report, the White House, OpenAI and Anthropic had not publicly responded. If the approach becomes formal policy, it would put national-security review directly into the international model-testing process — an area that has until now relied heavily on cooperation between labs and external safety organisations.

Source: CNA / Reuters report

Microsoft says a phishing-as-a-service platform called EvilTokens helped attackers compromise more than 12,000 inboxes across more than 10,000 organisations.

AI was not responsible for breaking into the accounts by itself. Instead, EvilTokens used AI to make an existing criminal workflow much more efficient: analysing compromised mailboxes, identifying relationships and payment authorities, and helping attackers craft believable fraud messages. That distinction makes the story more useful than generic claims about “AI hackers” — it shows how AI can compress the time and expertise needed after an account has already been breached.

Source: Microsoft Security

One of the clearest shifts is the gap between the cost of using AI and the cost of building everything around it.

Models are getting cheaper and more capable, but running them at scale still depends on expensive infrastructure, reliable power, strong security and a growing set of rules around how they can be used.

That creates an interesting tension. The technology itself is getting easier to access, while the systems needed to support it are getting harder and more expensive to build.

O T H E R__A I _N E W S

A few more releases, investments and developments worth having on your radar:

Google launched Gemini 3.8 Flash TTS and Flash-Lite TTS, adding prompt-designed voices, line-by-line performance control and voice replication from a 30-second sample, with built-in consent verification and watermarking.

Snorkel AI raised $350 million at a $3.5 billion valuation, nearly tripling its valuation as demand grows for specialised datasets and reinforcement-learning environments used to train advanced models.

Qualcomm agreed to acquire PickNik Robotics, the company behind the widely used open-source MoveIt robotics framework. Qualcomm says MoveIt will remain open source while becoming more closely integrated with its Dragonwing robotics platforms.

NVIDIA released Isaac ROS 5.0, adding agent-ready development workflows and expanding its open-source robotics stack across Jetson hardware. NVIDIA says the broader ROS ecosystem now reaches nearly 1.3 million users.

Google open-sourced AX, an Apache-2.0-licensed orchestration system designed specifically for long-running, stateful AI agents, with Kubernetes-style primitives for tasks, workspaces, models and network access.

Databricks made Genie One MCP generally available, giving AI tools and coding agents a governed route into company data and business context through the Model Context Protocol.

Amazon blocked Meta's Muse from shopping on Amazon.com, saying third-party agents should identify themselves and respect a service provider's decision about whether to participate. The dispute is an early test of who controls access when AI agents start shopping across other companies' platforms.

The UN's Independent International Scientific Panel on AI published a new brief on agent misalignment and loss-of-control risks, drawing on recent real-world testing incidents. The panel explicitly says it is not estimating the probability or timing of severe loss of control, but argues that current evidence cannot provide assurance that humans will retain control over more capable systems.

See you soon,

News Digest AI team