Lumina Digest

OpenAI Launches GPT-5.6 in Three Pricing Tiers and ChatGPT Work, the Office Agent

OpenAI brings frontier intelligence to three tiers (Luna, Terra, Sol) and turns ChatGPT into an agent that performs office work. But independent sources temper the leadership claim: on complex coding, Claude Fable 5 remains ahead.

OpenAI released GPT-5.6 on July 9, a three-tier family designed to route requests based on task difficulty: Luna ($1 input / $6 output per million tokens), Terra ($2.50 / $15), and Sol ($5 / $30). The price spread is 5x between the cheapest and the most capable. All models share a 1M-token context window, a maximum output of 128K, and a knowledge cutoff of February 16, 2026, as Simon Willison notes.

The technical novelty is programmatic tool calling: the model writes JavaScript executed in an isolated V8 runtime with no network access, so as to orchestrate multiple tool calls. OpenAI claims token reductions of 38% to 63.5% for cited customers; the feature is exposed in the Responses API, alongside a multi-agent beta. Caching changes too: cache writes cost 1.25x non-cached input, reads retain the 90% discount, with a minimum duration of 30 minutes.

The term "frontier," however, needs qualifying. On agentic tasks (Agents' Last Exam), Sol scores 53.6, ahead of Claude Fable 5 by 13.1 points. But on complex coding, Fable 5 remains ahead on SWE-Bench Pro (80% versus Sol's 64.6%). OpenAI counters that roughly 30% of those tasks are "broken." Willison, independently, observes that GPT-5.6 did not strike him as better than Fable on hard coding tasks: it competes mainly on cost per task, not in absolute terms.

In parallel, ChatGPT Work debuts, an agent that runs on GPT-5.6 and performs tasks across apps, files, and workflows, producing documents, spreadsheets, presentations, reports, and websites. The rollout starts Thursday on web and mobile, first for Pro, Enterprise, and Edu users, then Plus and Business.

Why it matters

• Entrepreneurs: The 5x tiered pricing lets you route tasks by cost, and ChatGPT Work can absorb duties currently handled by vertical tools (documents, reports, presentations). It forces a rethink of the AI budget and an assessment of the lock-in risk on a single agent-vendor.
• End users: ChatGPT stops being just a chatbot and becomes an agent that produces finished output on your behalf across apps and files. Convenient, but precisely because it accesses files and workflows with broad permissions, the results must be verified before use.
• LLM builders / devs · Frontier research: Programmatic tool calling and aggressive caching cut tokens and latency in multi-tool workflows, but the lead is not absolute: Sol leads on agentic tasks and trails Fable 5 on complex coding, with the same benchmarks contested (OpenAI: ~30% of SWE-Bench Pro tasks "broken"). Both model selection and the reading of frontier claims remain to be done benchmark by benchmark.

Grok 4.5: Aggressive Pricing ($2/$6) and Co-Training With Cursor, but Hallucinations Doubled and Already Surpassed by GPT-5.6 in Independent Benchmarks

SpaceXAI launches Grok 4.5 at $2/$6 per 1M tokens, co-trained with Cursor. In the Artificial Analysis index of July 8 it ranked fourth, behind Fable 5, GPT-5.5, and Opus 4.8. But on July 9 the new GPT-5.6 family surpassed it, threatening its price advantage as well; meanwhile the hallucination rate more than doubles.

SpaceXAI made Grok 4.5 available on July 8, 2026, ahead of the July 9 date that Musk had indicated as the public launch. It is a 1.5-trillion-parameter model — three times its predecessor — with a context window reduced from 1M to 500k tokens. The competitive lever is price: $2 per 1M input tokens and $6 for output, far cheaper than the flagships compared in the July 8 sources. Against GPT-5.5's $5/$30 and Fable 5's $10/$50, the estimated cost drops to $2.49 per task versus $5.07. The pricing lead is, however, already relative: on July 9 Artificial Analysis reports GPT-5.6 Luna at $1/$6 per 1M tokens, cheaper on input, with the new GPT-5.6 family redefining the cost/intelligence frontier. A caveat for the European market: at launch Grok 4.5 is not yet available in the EU, neither in products nor via API, with availability expected in mid-July.

Elon Musk presented it as "an Opus-class model, but faster, more token-efficient, and cheaper", specifying that it is "roughly comparable to Opus 4.7." The positioning, however, is not at the top of the independent benchmarks. In the July 8 Artificial Analysis analysis — before the July 9 GPT-5.6 results — Grok 4.5 ranked fourth in the Intelligence Index, with a score of 54, behind Fable 5, GPT-5.5, and Opus 4.8. The next day the picture changed again: on July 9 GPT-5.6 Sol (59) and Terra (55) surpassed Grok 4.5's score of 54. Musk's reference was in any case to Opus 4.7, the previous Opus generation and not the current flagship: a more modest comparison than the "Opus-class" label suggests.

On the benchmarks the picture is mixed: nearly on par on Terminal Bench 2.1 (83.3% versus Fable 5's 84.3%), but with wide gaps on DeepSWE 1.1 (53% versus 70%) and SWE Bench Pro (64.7% versus 80.4%). The most critical figure is calibration: on the AA-Omniscience index the hallucination rate more than doubles, from 25% to 54%, while accuracy rises from 35% to 52% — "it knows more, but is more confident when it is wrong."

The co-training with Cursor, which SpaceX agreed to acquire in mid-June for $60 billion in stock, brings a caveat about the transparency of the results, however: Cursor's benchmark was withdrawn after the discovery that an earlier snapshot of its code had accidentally ended up in the training data.

Why it matters

• Entrepreneurs: Frontier pricing collapses: near-frontier intelligence at a fraction of the cost ($2.49 per task versus $5.07) puts pressure on every provider's price list and lowers the threshold for embedding AI into products. But the advantage must be read over time: the 'Opus-class' positioning was already only fourth in the Artificial Analysis index of July 8, and since July 9 GPT-5.6 (Sol and Terra) has leapfrogged it, threatening its pricing lead as well with Luna at $1/$6. In choosing a vendor, what matters is up-to-date third-party benchmarks, not marketing; also worth weighing is EU availability, expected in mid-July.
• LLM builders / devs: An affordable, IDE-native model, co-trained with Cursor, is appealing for agentic coding loops. Before building on top of it, though, its coding results must be validated independently: the Cursor benchmark was withdrawn due to training-set contamination and the hallucination rate has more than doubled. Operational note for those developing in the EU: at launch, API access is not yet active.
• Frontier research: A textbook case of calibration degrading at scale: accuracy grows but hallucinations more than double (from 25% to 54%). The model 'knows more but is more confident when it is wrong' — a concrete signal about the reliability cost of scaling.

Claude Cowork Comes to Mobile and Web: Anthropic's Office Agent Steps Beyond Desktop-Only

Anthropic is extending its Cowork agent in beta to iPhone, iPad, Android, and the web, starting with the Max plan: tasks kicked off from the computer complete in the background and can be picked up from your phone. The desktop, however, remains the full experience.

On July 7, 2026, Anthropic extended Claude Cowork to iPhone, iPad, and Android (from the Claude app's sidebar) and to the web at claude.ai, as 9to5Mac reports. The agent had launched in January 2026 as a desktop-only app. Cowork is the space where you "hand Claude a task and it works across your files, calendar, email, messaging apps, the web, and the other connected tools until the job is done".

The cross-platform extension enables three capabilities: task continuity across devices, background execution without needing an active connection, and approval checkpoints. The latter kicks in before the agent wraps up. In practice, you start a task at your desk, receive updates on your phone, and pick up the finished output later, even with the laptop closed, as TechCrunch notes. The desktop, however, remains the full experience: according to Anthropic, only there can Cowork also use local files and the browser, while the new cloud sessions serve cross-device continuity and background work.

The rollout is in beta and gradual: it starts with the Max plan, with other plans in the following weeks; usage limits are extended through August 5, 2026.

On how Cowork is already being used, an Anthropic analysis of 1.2 million sessions (May 2026 data, more than 600,000 organizations) shows that over 90% of tasks are not software development: business operations at 33.4%, content creation and copywriting at 16.4%, software development at just 8.7% (TechCrunch, PYMNTS). TechCrunch frames the move as the "coding agent wars" spilling into the office: OpenAI, too, is pushing its own agent beyond development, toward reports, spreadsheets, and presentations. Two caveats to weigh: those statistics come from Anthropic's internal analysis, and the product remains in beta, with availability starting from the most expensive plan.

Why it matters

• End users: Cowork becomes usable on the go: you can delegate a task from the computer and collect it already done from your phone, even with the laptop closed, because the agent works in the background with approval checkpoints. Two limits remain, though: access starts with the Max plan, the most expensive one, and expanding to the other plans will take weeks; and the desktop remains the full experience — only there does Cowork use local files and the browser, while mobile and web serve to follow along and pick up the work.
• Entrepreneurs: With over 90% of tasks already non-coding and nearly half split between business operations and content production, the agent breaks out of the developer enclave and moves into everyday office operations: it's the concrete signal that delegating administrative and content tasks to an agent is maturing, in direct competition with OpenAI's 'agentic office' push.

Meta Deactivates the Feature That Generated AI Images by @-Mentioning Public Instagram Profiles; Muse Image Remains Available

All it took was @-mentioning an eligible public Instagram account in Meta AI to generate AI images from its posts, with no permission or notice. After criticism over deepfakes and likeness, Meta switched that feature off — but the Muse Image model stays active and the images already generated are still out there.

Meta has deactivated the Muse Image feature that let users generate images in Meta AI by @-mentioning public Instagram profiles; the Muse Image model — the first image generator from Meta Superintelligence Labs — remains available for its other functions (Meta). The feature, launched on July 7, 2026, let users @-mention the handle of an eligible public Instagram account inside a Meta AI prompt to create images derived from its posts. All of this without asking the account holder's permission and without any notification. It was not, however, open to just anyone: only public accounts of adult users were eligible, while private accounts and those of users under 18 were automatically excluded. The rollout, moreover, was initially limited to the United States (Forbes). For eligible accounts the feature was on by default (opt-out). To opt out, an adult with a public profile had to turn off the “Allow people to reuse your content on Instagram and with AI features at Meta” option (a toggle on Posts and Reels) or make their profile private (Malwarebytes).

The mechanism also exposed public figures to unauthorized deepfakes. The CAA agency — which represents Tom Hanks and Meryl Streep, among others — reaffirmed that “no name, image, voice, or creative work should be used by third parties, including AI models, without clear and documented consent,” while the SAG-AFTRA union urged its members to opt out. Within a few days Meta deactivated the feature: “We heard feedback that this feature missed the mark, so it's no longer available” (Engadget).

Two caveats remain material. The deactivation only blocks future generations: images created before the opt-out stay in circulation. And the bottom line doesn't change: Instagram head Adam Mosseri has argued that reliably detecting AI content is increasingly difficult. The path forward, in his view, is still to bet on labels and transparency — to the point of finding it more practical to certify authentic content than to flag fake content — rather than blocking AI entirely (Engadget).

Why it matters

• End users: Adults with a public Instagram profile in the United States were included in the feature by default, which let others generate AI images from their posts (private accounts and users under 18 were automatically excluded). It's still worth reviewing your content-reuse settings (or considering a private profile), because the opt-out only stops future generations and the images already created stay online.
• Entrepreneurs: It's a concrete risk for brands, corporate faces, and creators in partnerships: likeness and content exploitable without consent, with the reaction from CAA and SAG-AFTRA signaling legal and reputational exposure. The case also highlights the pattern to keep an eye on — platforms that launch AI features on an opt-out basis and disable them only under pressure — whenever you entrust data or images to third-party services.

A Chip That Mimics the Cerebellum: Diagnosing Arrhythmias at 98% With 10,000 Times Fewer Computations

At Northwestern, an asymmetric molybdenum disulfide memtransistor emulates the cerebellum: it ignores the routine and reacts to the anomaly. In ECG tests, it detects irregular heartbeats with over 98% accuracy while using a fraction of the operations of traditional AI.

A group at the McCormick School of Engineering of Northwestern University has built an electronic device that mimics the cerebellum, the brain center that filters out routine stimuli to focus resources on unexpected events. The work, led by Mark C. Hersam with Vinod K. Sangwan, Indira M. Raman, and Amit Trivedi, was published on July 10, 2026, in Nature Communications.

The component is an asymmetric molybdenum disulfide (MoS₂) memtransistor, a semiconductor just a few atoms thick, in which one electrode partially overlaps the semiconductor through a thin insulating layer. When the direction of the voltage is reversed, the device switches between an excitatory and an inhibitory mode, replicating the balance of cerebellar circuits: it stays stable during normal activity and reacts only when something changes. "The cerebellum is excellent at ignoring the expected and reserving its resources to react to the unexpected," Hersam explains.

In electrocardiogram tests, the device detected abnormal heartbeats with over 98% accuracy, within one-fifth of a heartbeat and at more than twice the speed of conventional AI. And it does so using roughly 10,000 times fewer computational operations. The result extends a line of work from 2023 (Nature Electronics), in which two memtransistors performed classifications that would have required more than 100 traditional transistors, at one-hundredth of the energy: the novelty here is applying that paradigm to cerebellum-inspired anomaly detection rather than to the cortex. Among the envisioned applications are health wearables, medical monitoring, autonomous vehicles and robots, and cybersecurity.

Explicit limitations remain: the team emulated "only part of the cerebellum's neural circuit", without yet any learning and adaptation capability, and the figures come from a laboratory demonstration (funded by the NSF), not from a field deployment or an independent replication.

Why it matters

• Frontier research: It points to a concrete direction for bio-inspired neuromorphic computing: shifting the reference model from the cortex to the cerebellum for anomaly detection offers a hardware alternative to traditional neural networks, with energy efficiency orders of magnitude higher. It remains, however, a laboratory demonstration that emulates only part of the circuit and still needs to be validated, scaled, and independently replicated.
• End users: It foreshadows wearables and health-monitoring devices capable of running advanced diagnostics locally — such as arrhythmia detection — with minimal power consumption and longer battery life, keeping sensitive health data on the device instead of in the cloud.