GPT-5.6 Sol Terra Luna in preview, Anthropic Economic Index, Gemma 4 on Cerebras: AI watch for June 26, 2026

June 26, 2026 is a packed day on several fronts: OpenAI unveils the limited preview of the GPT-5.6 series (Sol, Terra, Luna), a new naming system with lower prices on Terra and Luna; Anthropic publishes its sixth Economic Index, “Cadences,” with the results of a survey of 9,700 users; Gemma 4 crosses 200 million downloads in 2.5 months and becomes the first multimodal model on Cerebras at 1,500 tokens per second. On the developer tooling side, GitHub Desktop 3.6 adds worktrees and the Copilot SDK, Copilot for Jira reaches general availability, and the LSP Setup skill brings semantic intelligence to Copilot CLI across 14 languages. HeyGen announces $200 million in ARR in eight months, and Cohere positions itself around enterprise sovereignty in response to government restrictions on GPT-5.6.

GPT-5.6 Sol, Terra and Luna — limited preview of the new OpenAI generation

June 26 — OpenAI launches a limited preview of the GPT-5.6 series, made up of three models with distinct positioning. The series introduces a new naming system: the number (5.6) identifies the generation, while Sol, Terra, and Luna designate durable capability tiers that can evolve independently.

“Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work.” — @OpenAI on X

Three models, three capability levels

GPT-5.6 Sol: next-generation flagship model, the most capable to date for cybersecurity, code, and biology. Introduces a maximum reasoning effort level and an ultra mode that uses sub-agents to speed up complex tasks.
GPT-5.6 Terra: balanced model for everyday work, with performance comparable to GPT-5.5 but twice as cheap.
GPT-5.6 Luna: fast and economical model for high volumes, at the lowest cost in the range.

Pricing and performance

Model	Input (per million tokens)	Output (per million tokens)	Positioning
GPT-5.6 Sol	$5	$30	Flagship / Frontier
GPT-5.6 Terra	$2.50	$15	Balanced, 2× cheaper than GPT-5.5
GPT-5.6 Luna	$1	$6	Fast / economical

On benchmarks, Sol sets a new state of the art on Terminal-Bench 2.1 (code/CLI) and outperforms GPT-5.5 on GeneBench v1 (biology) with fewer tokens. On ExploitBench² (cybersecurity), Sol is comparable to Mythos Preview with about one-third of the output tokens. Sol will be available on Cerebras at up to 750 tokens per second starting in July 2026.

Strengthened safety and restricted availability

OpenAI says it has deployed the most robust safety stack to date: more than 700,000 A100-equivalent GPU hours devoted to automated red teaming, real-time classifiers for cybersecurity and biology, and account-level review to detect persistent malicious behavior. GPT-5.6 Sol does not cross the “Cyber Critical” threshold in OpenAI’s Preparedness Framework.

At the request of the U.S. government, OpenAI is starting with a limited preview for a small group of trusted partners, before a broader rollout “in the coming weeks” via the API, Codex, and ChatGPT.

🔗 OpenAI Blog — GPT-5.6 Preview

Anthropic Economic Index — sixth report “Cadences”

June 26 — Anthropic publishes the sixth report of the Anthropic Economic Index, titled “Cadences.” It is the first edition to combine continuous hourly sampling, a classifier for produced artifacts, and the results of a survey of 9,700 Claude users.

New methodology

Change	Description
Hourly sampling	Data collected continuously (no longer on a 7-day window) — observable hour-by-hour variations
Artifact classifier	Each conversation categorized by its main output (explanation, document, code, etc.)
Granular data	Split between Claude conversations (chat + Cowork) and 1P API, aggregated at the monthly level

The rhythms of usage

Usage reflects the calendar and daily habits: the share of personal conversations rises from ~35% on weekdays to ~50% on weekends. Cooking recipes are requested 2.3 times more often at 6 p.m. than on average; sleep advice peaks around 5 a.m. Tax-related requests were eight times more frequent on April 14 (the day before the U.S. filing deadline) than the May average. On Claude Code, the uses that drop the most on weekends are back-end architecture and API debugging; the ones that rise the most: AI agent design, quantitative trading, and video games.

What Claude produces

93% of conversations produce an identifiable artifact: explanations (17%), documents and reports (15%), advice (11%), code and scripts (~16% combined). Apps and websites consume more than three times the tokens of a median conversation. Claude responds, on average, at a reading level about one year of schooling above the prompt — the gap is largest for images and charts (+2.6 years), games (+1.9), and apps (+1.7).

The survey of 9,700 respondents

Indicator	Value
Expect significant changes to their responsibilities within 12 months	>1/3
Think AI will be able to handle most of their tasks within 12 months	>35%
Report productivity gains in speed	86%
Report productivity gains in amount of work	82%
Report productivity gains in quality	69%
Learn more thanks to AI	68%
Feel AI increases the value of their skills	57%
Think it is likely they will lose their own job	~10%
Think it is likely their younger colleague will lose their job	>1/3

Core finding: the users who delegate the most to Claude are also the most optimistic about AI’s impact on their job and skills — without that being accompanied by a reduction in perceived learning. Women make up 12% of the paired sample; after controlling for differences in occupation, they use Claude Code 6.3 percentage points less, but spend more active time on chat.

🔗 Anthropic Economic Index — Cadences report

Claude Code v2.1.193 — autoMode, OpenTelemetry and bash autocompletion

June 25 — Claude Code v2.1.193, released late in the evening, brings several new features focused on automatic mode control, observability, and terminal ergonomics.

Feature	Detail
`autoMode.classifyAllShell`	All Bash/PowerShell commands go through the automatic mode classifier
Auto-mode refusal reasons	Recorded in the transcript, the toast, and the `/permissions` tab
OpenTelemetry `claude_code.assistant_response`	Logs the model response text (hidden by default; `OTEL_LOG_ASSISTANT_RESPONSES=1`)
Bash path autocompletion (`!`)	Suggested paths in real time as you type
MCP auth notification	Notice to `/mcp` if servers require authentication at startup
Idle shell auto-response in the background	Commands waiting for input receive an automatic response (disable via `CLAUDE_CODE_DISABLE_BG_SHELL_PRESSURE_REAP=1`)

Several regressions are fixed: pinned background agents are no longer re-prompted to “continue where you left off” after each automatic update, and the phantom “general-purpose (resumed)” sub-agent that restarted the main conversation is removed. The headersHelper MCP restarts and reconnects automatically after a 401/403 return.

🔗 Claude Code changelog

Google / Gemini — Gemma 4 on Cerebras, 200M downloads and June Drops

Gemma 4 × Cerebras — first multimodal model at 1,500 tokens/s

June 26 — Gemma 4 becomes the first multimodal model available on Cerebras. The 31B version runs at 1,500 tokens per second on Cerebras infrastructure. A 24-hour virtual hackathon is scheduled for June 28 with $5,000 in prizes for participants, who get early access to Gemma 4 on Cerebras via Luma (Cerebras × Google Gemma 4: $5,000 Hackathon).

🔗 @googlegemma tweet — Cerebras × Gemma 4

Gemma 4 — 200 million downloads in 2.5 months

June 25 — The Gemma 4 family surpasses 200 million downloads in just 2.5 months since launch. For context, the entire Gemma family totaled 100 million downloads at the launch of Gemma 3 — so Gemma 4 doubled that total in 2.5 months. This milestone was retweeted by @GoogleDeepMind, showing how much weight Google assigns to this result.

🔗 @googlegemma tweet — 200M downloads

Gemini Drops June 2026 — real-time voice images and SMB tools

June 26 — The Gemini app publishes its June monthly recap (“Gemini Drops”): real-time image generation with voice and new tools for small businesses. The tweet links to a Google blog post detailing the month’s features.

🔗 @GeminiApp tweet — Gemini Drops June 2026

GitHub Copilot — Desktop 3.6, Jira GA and semantic CLI intelligence

GitHub Desktop 3.6 — Worktrees and Copilot SDK

June 26 — GitHub Desktop 3.6 brings two major changes: support for Git worktrees and deeper Copilot integration through the new Copilot SDK.

Feature	Detail
Copilot SDK	Common foundation for all Copilot features in Desktop
Commit authoring	Generates messages respecting `.github/copilot-instructions.md` and `AGENTS.md`
Conflict resolution	AI explains conflicting changes and suggests a resolution to accept or edit
Git worktrees	Work on multiple branches in parallel without stashing or repeated cloning
Model picker	Choose the model for each Copilot feature
BYOK	Connect to a third-party provider or a local model

Available for macOS and Windows (GitHub Desktop 3.6.0). Worktrees are especially useful with coding agents that create isolated workspaces (worktrees).

🔗 GitHub Changelog — Desktop 3.6

GitHub Copilot for Jira — General Availability

June 25 — GitHub Copilot for Jira moves into general availability (GA) after a public preview since March 2026. GA brings two new capabilities: progress streaming (coding agent updates sent in real time to the Jira ticket without switching context) and post-session steering (after opening a draft PR, the user can give new instructions directly in Jira; the agent continues on the same PR). Organization onboarding for GitHub and repositories is simplified.

🔗 GitHub Changelog — Copilot for Jira GA

LSP Setup skill — semantic intelligence across 14 languages for Copilot CLI

June 25 — GitHub Copilot CLI gets a new LSP Setup skill that gives it genuine semantic code intelligence — definitions, references, types — via LSP (Language Server Protocol) servers, replacing the grep/decompilation approach. Coverage: 14 languages. LSP server configuration and installation are handled by the skill.

🔗 GitHub Blog — LSP Setup skill · @github tweet

Open models and research — HF Jobs vLLM and Sakana CoffeeBench

HF Jobs — OpenAI-compatible vLLM server in one command

June 26 — Hugging Face publishes an official guide for launching a private, OpenAI-compatible vLLM server on HF Jobs infrastructure with a single command (hf jobs run --flavor a10g-large --expose 8000 vllm/vllm-openai:latest vllm serve Qwen/Qwen3-4B). No Kubernetes, no servers to provision, billed by the second.

Aspect	Detail
Available GPUs	A10g-large (1.50 USD/h), H200×2, H200×8
Protocol	OpenAI-compatible (ChatCompletions), HF token authentication
SSH access	`hf jobs run --ssh` + `hf jobs ssh <job_id>`
Tested models	Qwen3-4B, Qwen3.5-122B-A10B (MoE, 2×H200)
Gradio interface	Streaming support with collapsible reasoning
Coding agent	Pi framework integration (provider-agnostic)
Inference Endpoints difference	Jobs = experimentation/evals, Endpoints = production with scale-to-zero

Prerequisite: huggingface_hub >= 1.20.0.

🔗 HF Blog — vLLM on HF Jobs

Sakana CoffeeBench — 90-day multi-agent simulation of the coffee supply chain

June 26 — Sakana AI, in collaboration with Azusa Audit Corporation, publishes CoffeeBench: a benchmark evaluating the long-term management capabilities of LLM agents. Six companies (farmers, roasters, retailers), each managed by a distinct agent, conduct price negotiations, place orders, and manage their inventories over 90 simulated days with the goal of maximizing net profits. The results reveal significant behavioral differences between models: some agents negotiate actively, others analyze without acting and end up in the red. The research will be presented at the ICML 2026 workshop “Failure Modes in Agentic AI.” The tweet received 173,800 views.

🔗 Sakana — CoffeeBench · arXiv 2606.16613 · @SakanaAILabs on X

HeyGen surpasses 200 million USD in ARR in eight months

June 25 — HeyGen, the identity-first AI video platform, announces that it has surpassed 200 million USD in ARR (annual recurring revenue), doubling in eight months. The growth reflects rapid adoption both by independent creators and by 85% of Fortune 100 companies.

Metric	Value
ARR	200 million USD (doubled in 8 months)
Users	30 million
Countries	196
Languages	175+
Fortune 100	85%
Videos created	118 million
Capital efficiency	2.70 USD ARR per dollar raised

HeyGen highlights Avatar V (15 seconds of video are enough to create a digital twin), HyperFrames (open source Apache 2.0 framework for agentic video creation, 21,600 GitHub stars in one month), and HeyGen MCP integration for AI agent workflows. 63 features were launched in 3 months.

🔗 HeyGen Blog — 200 million USD ARR

Cohere — enterprise sovereignty in the face of GPT-5.6 restrictions

June 25 — Cohere publishes an aggressive positioning tweet on sovereignty and enterprise control, in reaction to OpenAI’s announcement limiting GPT-5.6 to a small group of partners at the request of the U.S. government.

“When you use Cohere, there are no staggered releases. No sudden disablements. We trust you completely: ‘[The customer] is in full control. We can’t see in, we can’t switch it off.’” — @cohere on X, citing Aidan Gomez, CEO of Cohere

This positioning comes on the same day OpenAI announces that GPT-5.6 Sol, Terra, and Luna will be available “generally only in the coming weeks,” for now only to a small group at the request of the U.S. government. The tweet received 18,900 views.

Briefs

Warp — Cloud Software Factory series (article 1) (June 25) — First article in a Warp series on building a cloud software factory with skills and loops. This episode covers the “automatic triage skill” for automating ticket triage in a development workflow. Educational content (7 min read). 🔗 Warp Blog
Sakana Fugu — arXiv technical report 2606.21228 (June 26) — Official Fugu technical report published on arXiv, complementing the June 22 launch. Documents the Fugu, Fugu-Ultra, and Fugu-Llama architectures on coding, reasoning, roleplay, and agentic task benchmarks (including SWE-Bench Pro). 🔗 arXiv 2606.21228 · Tweet @SakanaAILabs
Antigravity 2.2.1 (June 25) — Google’s AI development environment releases version 2.2.1: new integrated “Antigravity Guide” skill, audio file rendering, substring file search, 19 improvements, and 17 fixes. Previous version: 2.1.4 from June 11. 🔗 Antigravity — changelog
Copilot code review — efficiency updates (June 25) — Copilot code review now uses native CLI/SDK tools (grep, rg, glob, view), reducing costs by 20% while maintaining quality. Medium analysis depth enables attribution in the PR summary comment and organization-level review depth configuration. 🔗 GitHub Changelog
MAI-Code-1-Flash extended to Copilot Business and Enterprise (June 26) — The MAI-Code-1-Flash model is now available to Copilot Business and Copilot Enterprise plans. 🔗 GitHub Changelog
Cohere — “Taps sign” meme reaction (June 26) — In response to restricted access to GPT-5.6 Sol, Cohere posts a humorous meme on X featuring a photo of Aidan Gomez on the cover of FORTUNE, implying their CEO has already said the essential point. The tweet reaches 37,800 views and 444 likes. 🔗 Tweet @cohere

What it means

Fragmentation of access to frontier models is becoming a business issue. GPT-5.6 Sol is launched in limited preview at the request of the U.S. government — an unprecedented phased rollout at this scale for OpenAI. Cohere responds immediately with a position on full customer sovereignty (“We can’t see in, we can’t switch it off”), signaling that enterprise control over AI models is becoming a differentiating sales argument. The following day’s meme confirms that this positioning is deliberate and coordinated. The GPT-5.6 Sol/Terra/Luna series also represents a notable naming shift: durable names that can evolve independently, rather than a fixed version number.

Measuring AI’s economic impact is gaining depth and nuance. Anthropic Economic Index “Cadences” is the first edition to combine real usage data with a survey of 9,700 respondents. The results add nuance to the dominant narrative: 86% report productivity gains in speed, but the users who delegate the most are also the most optimistic — without this being accompanied by a reduction in perceived learning. The gender gap (12% women in the paired sample, -6.3 points of Claude Code usage) remains poorly explained but is documented for the first time at this granularity. The temporal rhythms of usage (tax filings at 2 p.m., earnings at 6 p.m., sleep at 5 a.m.) offer a concrete reading of how AI fits into everyday life.

Development team tooling is densifying simultaneously on several fronts. GitHub Desktop 3.6, Copilot for Jira GA, and the LSP Setup skill for Copilot CLI arrive on the same day, forming full coverage of the development cycle: desktop (Desktop), project management (Jira), terminal (semantic CLI on 14 languages). Claude Code v2.1.193 adds OpenTelemetry observability for model responses and automatic shell command classification, two features that address enterprise compliance and traceability needs. HF Jobs vLLM, meanwhile, lowers the entry barrier for on-demand inference without infrastructure to manage.

Open models and AI video are reaching commercial milestones that reshuffle the board. Gemma 4 doubles in 2.5 months the historical total of the Gemma family at the launch of Gemma 3, and becomes the first multimodal model to run on Cerebras at 1,500 tokens/s. HeyGen (200 million USD ARR, 2.70 USD ARR per dollar raised) shows that AI video can achieve rare capital efficiency in the tech sector. Sakana’s CoffeeBench illustrates a third form of contribution: multi-agent benchmarks over long horizons (90-day supply chain) that capture behaviors classical evaluations do not see — gaps that will be presented at ICML 2026.