The AI Dispatch — May 8, 2026

Revenue

Anthropic Hits $30B Revenue Run Rate, Signs $1.8B Akamai Compute Deal

At the company’s Code with Claude developer conference, CEO Dario Amodei disclosed an 80× annualized growth surge in a single quarter — and announced the company’s third nine-figure-or-larger compute contract in as many weeks.

By AI Dispatch Desk · May 8, 2026 · Sources: VentureBeat · Bloomberg

Anthropic disclosed on Friday that its annualized revenue had surged from roughly $9 billion at the end of 2025 to approximately $30 billion by April 2026 — an 80× increase on an annualized basis in a single quarter. CEO Dario Amodei delivered the figures during the keynote of the company’s Code with Claude developer conference. Claude Code, Anthropic’s coding agent platform, alone accounts for roughly $2.5 billion of that annualized revenue. More than 1,000 enterprise customers now pay Anthropic over $1 million per year. The disclosure represents the most aggressive single-quarter growth figure ever publicly reported by a frontier AI lab.

On the same day, Anthropic signed a seven-year, $1.8 billion cloud computing contract with Akamai Technologies — the largest deal in Akamai’s history, according to the company. Unlike Anthropic’s recent compute deals with Google and SpaceX/Colossus, which target training capacity, the Akamai contract targets inference. The arrangement gives Anthropic access to Akamai’s 4,200-location edge network — a distributed footprint built for low-latency content delivery that the two companies are now repurposing for low-latency model serving. Akamai shares closed up 27% on the announcement, the largest single-day gain in the company’s history.

The Akamai deal is the third major compute commitment Anthropic has disclosed in a span of weeks. The Google arrangement, valued at $40 billion in cumulative cloud spend, anchors the company’s training capacity through 2030. The SpaceX/Colossus deal, valued at approximately $15 billion per year, brings the Memphis-area xAI-affiliated supercomputer cluster online for Anthropic workloads. The Akamai deal closes the loop by adding distributed inference. Taken together, the three contracts represent a capacity commitment that is unprecedented in scale for a single AI customer — and that is now backed, the revenue disclosure makes clear, by demand growth on a comparable scale.

The $30 billion run rate marks Anthropic’s decisive arrival at a revenue tier that until recently only OpenAI had reached. Amodei’s keynote framed the growth as the consequence of three converging trends: enterprise adoption of Claude as a default coding assistant, the expansion of Claude Code into autonomous multi-step workflows, and the shift in Fortune 500 procurement toward seven-figure annual contracts as the standard frontier-model commitment. The 1,000-plus enterprise accounts paying over $1M annually is a figure with no public analog at OpenAI, which has historically declined to disclose customer concentration metrics at that resolution.

What the numbers do not yet show is unit economics. Anthropic does not break out gross margin on inference workloads, and the seven-year structure of the Akamai contract suggests the company is locking in capacity pricing rather than betting on commodity declines. The $2.5 billion Claude Code line, in particular, will draw scrutiny: coding agents are token-hungry, and the workloads that produced the growth are also the workloads most exposed to margin compression if competitors match capability at lower prices. None of that erases the headline figure. An 80× annualized increase in a quarter is a number the field has never seen before from any AI company; it forces a re-pricing of every assumption about where the frontier-lab revenue curve flattens.

OpenAI

OpenAI Opens Self-Serve ChatGPT Ads Manager and Ships GPT-Realtime-2 Voice

By AI Dispatch Desk · May 8, 2026 · Sources: openai.com/new-ways-to-buy-chatgpt-ads · openai.com/advancing-voice-intelligence

OpenAI executed two product launches the same day — one commercial, one technical — that together reframe how the company expects to monetize the ChatGPT surface and how it expects the API economy around it to evolve. The Ads Manager opens self-serve ad buying on ChatGPT to anyone with a credit card and a Stripe account. The Realtime API release ships three new voice models, including a GPT-5-class reasoning variant that the company is positioning as the first voice model capable of multi-tool agentic loops without an intermediate text representation.

The new self-serve ChatGPT Ads Manager portal eliminates the prior $50,000 minimum spend floor that had previously gated participation in the ads pilot to large brands and agencies. The portal supports both CPC and CPM bidding, includes the standard set of targeting controls (geography, interest, behavioral signals), and offers automated campaign optimization driven by the same recommendation models that power ChatGPT’s organic ranking. The ads pilot, which has been running in the U.S. and Australia, will expand in the coming weeks to the United Kingdom, Mexico, Brazil, Japan, and South Korea. OpenAI’s stated internal targets, reaffirmed in the launch post, are $2.5 billion in advertising revenue in 2026 and $100 billion annually by 2030 — figures that, if hit, would make ChatGPT advertising one of the largest greenfield ad markets created in a decade.

The Realtime API release is the technical companion piece. OpenAI shipped three new voice models: GPT-Realtime-2, a GPT-5-class reasoning model with a 128K-token context window and multi-tool calling support, capable of streaming voice in and voice out while invoking external tools mid-utterance; GPT-Realtime-Translate, a live-translation variant accepting input in 70+ languages and producing output in 13; and GPT-Realtime-Whisper, a streaming speech-to-text model designed for low-latency transcription workloads. The Realtime-2 model in particular eliminates a longstanding architectural limitation: previous voice models had to round-trip through text to invoke tools, adding latency and breaking the natural cadence of conversation. Realtime-2 invokes tools directly from the voice stream, allowing for genuinely interactive agentic voice applications — a capability that, alongside the Ads Manager, signals the surface area OpenAI now expects developers to build against.

The two announcements have a shared subtext. Self-serve ads put ChatGPT in the same revenue-collection posture as Google Search and Meta’s social properties — an at-scale, programmatic, low-friction monetization layer that runs without dedicated sales staffing. Realtime-2 broadens the surface beyond text-and-image chat into voice-native applications, where the addressable market includes call centers, accessibility tools, in-car interfaces, and live consumer voice agents. Together they signal that OpenAI’s 2026 monetization strategy is no longer about scaling enterprise API contracts alone; it is about owning two distinct flywheels — consumer ad revenue, and developer-built voice applications — that compound on top of the ChatGPT subscription base.

Open Weights & Privacy

The Friday Wire

A MIT-licensed 8B image model takes the top open-weight diffusion slot, while Meta retires Instagram’s opt-in encryption and tells users to switch apps for private chats.

Open Weights

HiDream-O1-Image (8B) Open-Sourced Under MIT, Tops Open Diffusion Leaderboard

Source: huggingface.co/HiDream-ai

HiDream-ai open-sourced HiDream-O1-Image on Friday under the MIT license, releasing an 8B-parameter image generation and editing model along with undistilled and distilled Dev variants and a Reasoning-Driven Prompt Agent. The architecture is unusual: the model is built on a Pixel-level Unified Transformer (UiT) that operates without an external variational autoencoder, natively encoding raw pixels, text, and task conditions in a single shared token space. The unified design lets the same checkpoint handle text-to-image generation, image editing, and subject-driven personalization at output resolutions up to 2,048×2,048 without architecture-specific adapters. The Dev variant debuted at #8 on the Artificial Analysis Text-to-Image Arena leaderboard — the top spot among open-weight diffusion models, ahead of every prior open-source release in the category. The MIT licensing is the more consequential choice: until this release, the strongest open image models had shipped under non-commercial or research-only licenses, and the production-grade option had to be proprietary. HiDream removes that gap.

Privacy

Meta Kills Optional End-to-End Encryption on Instagram DMs

Source: MacRumors

Meta removed the opt-in end-to-end encryption feature from Instagram’s direct messaging on May 8, citing “very low uptake” and redirecting users who want encrypted conversations to WhatsApp. The change applies to text, images, and voice notes sent through Instagram DMs — all of which are now technically accessible to Meta on the company’s servers and, by extension, to law enforcement requests served on Meta directly. Privacy advocates immediately disputed the rationale, noting that opt-in E2EE was launched without significant promotion or default activation, which made the “low uptake” result close to predetermined. The deprecation comes amid sustained pressure from law-enforcement agencies in multiple jurisdictions for Meta to weaken or remove encryption across its platform set; the company has resisted those requests on WhatsApp but has now conceded on the surface where it had introduced the protections most recently. Users who valued the feature have, in practical terms, been offered a different app rather than a fix.

Research & Policy

From the Statehouses, From the Citation Trail

A 111-million-reference audit pins down the LLM-era hallucinated-citation problem; a transparency tracker counts 78 active chatbot bills across 27 states.

Research

Citation Audit Finds 147K Hallucinated References in 2025 Scientific Papers

Source: arXiv:2605.07723

Researchers audited 111 million references across 2.5 million papers from arXiv, bioRxiv, SSRN, and PubMed Central, looking for citations to works that do not exist. The post-LLM spike is unmistakable: the rate of non-existent references in the corpus jumps in 2023, climbs again in 2024, and plateaus at an elevated level in 2025. Conservative estimate for 2025 alone: 146,932 hallucinated references across the audited venues. The errors are not uniformly distributed. They concentrate in fields with rapid LLM uptake, in manuscripts that show other signs of AI assistance, and in small or early-career author teams — the cohorts least likely to have the senior co-author bandwidth required to verify every reference manually. A more troubling pattern: the hallucinated citations themselves skew credit toward already-prominent, predominantly male scholars, because LLMs reach for the most frequently-co-occurring names in a topic when asked to invent a citation. The result is a quiet, mechanical reinforcement of the field’s existing prestige distribution — through references to papers that were never written.

State Laws

Transparency Coalition Counts 78 Active Chatbot Bills Across 27 States

Source: Transparency Coalition

The Transparency Coalition’s May 8 legislative update catalogs 78 active chatbot safety bills across 27 U.S. states — the most concentrated state-level AI legislative activity in U.S. history, according to the group’s tracking. Utah’s session closes with nine AI-related bills sent to the governor’s desk, the largest single-state package this cycle. Colorado is advancing two distinct bills toward adjournment: a general-purpose chatbot safety act and a separate, narrower bill targeting AI systems marketed as mental-health or therapy companions. The combined pace of state activity is now running well ahead of any plausible federal trajectory; the White House has signaled interest in federal preemption legislation that would override the state patchwork, but no such bill has emerged from committee. Until preemption arrives, the operating compliance reality for any chatbot product with national distribution is a 27-state matrix that varies on definitions, disclosure thresholds, age gating, and liability standards.

Toolbox

Codex CLI v0.130.0 Ships Headless Remote-Control, Plugin Sharing, Bedrock Console Login

Source: github.com/openai/codex releases

OpenAI’s Codex CLI v0.130.0, released Friday, is the second workflow-dense version drop in two days and pushes the tool further toward embedded-agent and team-orchestration use cases. The headline change is a new headless entry point that lets external pipelines drive Codex as an agent without a UI; alongside it are formalized plugin sharing primitives, app-server thread pagination, and an expanded set of AWS Bedrock auth pathways.

codex remote-control: a new headless app-server entry point that lets external pipelines steer Codex as an embedded agent without a TUI. Designed for CI runners, evaluation harnesses, and multi-agent orchestrators that need to send turns and receive structured responses programmatically.
Plugin sharing: link metadata and discoverability controls let users share or unshare individual plugins or entire workspaces. Source filtering allows scoping shared plugins by repository or organization, and local share-path tracking surfaces what is shared from each machine.
App-server thread pagination with variable detail levels: thread replay for multi-agent sessions can now be paginated by chunk size and rendered at configurable fidelity, easing review of long agent histories without forcing a full-resolution load.
AWS Bedrock authentication now supports console-login credentials alongside the existing AWS profile and access-key flows, removing one of the most-cited setup friction points for Bedrock-routed Codex deployments.
Sandbox security hardening on both Linux and Windows, including tightened syscall filtering and improved isolation of plugin-spawned subprocesses. No specific CVE references in the release notes; the changes appear preventive rather than responsive.

The remote-control entry point is the change with the largest downstream surface. It makes Codex usable as a backend agent inside larger systems — CI pipelines, evaluation grids, multi-agent orchestrators — where the TUI has been the structural blocker. Paired with the plugin sharing primitives shipped the same day, the release reads as a deliberate move to position Codex as the agent runtime inside organizations that have already built tooling around it, rather than only the front-end terminal experience for individual developers.

Briefs

From the Desk

Three arXiv preprints worth a closer look: an endpoint-level inference benchmark, a six-tier open-weight agent leaderboard, and a vision-language reasoning scheme for long-horizon manipulation.

Token Arena

arXiv:2605.00300 reframes inference benchmarking at the endpoint level rather than the model level — testing 78 endpoints across 12 model families. The headline finding: the same model served from different endpoints can differ by up to 12.5 accuracy points on math and code benchmarks, depending on quantization, batching policy, decoding parameters, and tool-routing configuration. The implication for procurement is sharp: benchmark scores attached to a model name are only meaningful when paired with a specific endpoint configuration. The same OSS weight can be a top performer on one provider and a middle-of-the-pack performer on another — not because the weights changed, but because everything around them did.

Source: arXiv:2605.00300 · Token Arena

AgentFloor

arXiv:2605.00334 introduces a 30-task, six-tier benchmark for autonomous agents, evaluated across 16 open-weight models plus GPT-5 over 16,542 total runs. The benchmark spans tiers from single-step tool calls up through multi-day project workflows. Headline finding: the strongest open-weight model in the suite matches GPT-5 on aggregate score across all six tiers — though the per-tier breakdown reveals open-weight strengths concentrated in lower tiers and a widening gap as task horizon increases. The result is consistent with the broader 2026 picture: open-weight models have closed most of the gap on bounded tasks, while a meaningful frontier-only advantage persists on the long-horizon planning frontier.

Source: arXiv:2605.00334 · AgentFloor

Interleaved VLR

arXiv:2605.00438 introduces Interleaved Vision-Language Reasoning, a scheme that alternates textual sub-goals with visual keyframes during long-horizon manipulation tasks. Rather than reasoning entirely in text and then dispatching motor commands, the model emits a planned visual keyframe at each sub-goal boundary, conditioning the next reasoning step on a rendered prediction of the intermediate state. The interleaved variant beats single-modality baselines across the benchmark suite, with the largest margins on tasks requiring more than three sub-goals. The contribution is methodological more than architectural: the same underlying VLM is being used, but with a different prompting and rollout structure that better matches how long-horizon physical tasks decompose.

Source: arXiv:2605.00438 · Interleaved VLR

GitHub Trending — Friday Snapshot

Top repositories trending on May 8, 2026 — agent skills and harness collections lead the week.

GitHub Trending — Friday Snapshot
Repo	Language	Today’s Signal	What it does
mattpocock/skills	TypeScript	+1.6K this week (#1)	Reusable agent skills for AI coding workflows — the canonical Claude Code skills collection.
affaan-m/everything-claude-code	Shell / Markdown	~100K stars (#4 weekly)	30 agents, 136 skills, and 60 commands bundled as a complete Claude Code harness.
microsoft/typescript-go	Go	~25.5K stars	Native TypeScript compiler ported to Go — 5–10× faster builds in early benchmarks.
datawhalechina/easy-vibe	Markdown	~11.4K stars	Vibe-coding beginner programming course — learning materials for the LLM-assisted-coding cohort.
pingcap/ossinsight	TypeScript	Established	Open-source GitHub analytics powered by TiDB — the reference dashboard for OSS-trend tracking.