AI News Today: GPT-5.6 Tops Coding Benchmarks & Gemini 3.5 Pro Delayed – July 15, 2026

⚡ Today's AI Briefing (TL;DR)

GPT-5.6 launches: OpenAI's three-tier model family (Sol, Terra, Luna) claims the top spot on the Artificial Analysis Coding Agent Index, beating Claude Fable 5.
GLM-5.2 goes MIT-licensed: Zhipu AI's free, open-weight model becomes the strongest open coding model, nearly matching Claude Opus 4.8.
Gemini 3.5 Pro delayed to July 17: Google scrapped the 2.5 Pro architecture for a ground-up rebuild with a 2-million-token context window.
Claude Fable 5 is back: Anthropic redeployed its flagship model after export controls tied to a jailbreak exploit were lifted.

AI News (Top Updates)

1. OpenAI launches GPT-5.6 in three tiers, claims the coding benchmark crown

On July 9, OpenAI released GPT-5.6 as three variants — Sol (flagship, "best coding model yet"), Terra (mid-tier), and Luna (budget). On the Artificial Analysis Coding Agent Index, Sol scored 80, beating Anthropic's Claude Fable 5 by 2.8 points while using under half the output tokens and running roughly a third cheaper. OpenAI also shipped ChatGPT Work, a workplace-companion agent for drafting documents, spreadsheets, and slides.

Simple version: OpenAI's newest AI model is faster, cheaper, and (by its own benchmark) better at coding than Anthropic's best model.

2. Zhipu AI's GLM-5.2 becomes the strongest open-weight coding model — MIT licensed

Released with weights on Hugging Face and ModelScope, GLM-5.2 posts 74.4% on FrontierSWE — just a point behind Claude Opus 4.8 — and tops Artificial Analysis's Intelligence Index among open-weight models, all while running at a fraction of the cost of closed rivals.

Simple version: A free, downloadable Chinese AI model is now nearly as good at coding as the best paid models.

3. MiniMax-M3 tops Artificial Analysis's Intelligence Index for its price class

The Chinese lab's new reasoning model ranks #1 of 217 models tracked at its cost tier, with a 1-million-token context window and text/image/video input support, reportedly eclipsing GPT-5.5 and Gemini 3.1 Pro on several benchmarks at 5–10% of the cost.

Simple version: This model scores near the top of independent rankings while costing a fraction of what rivals charge.

4. Moonshot AI ships Kimi K2.7-Code, a 1-trillion-parameter open coding model

The update reports a 21.8% jump over K2.6 on Kimi Code Bench v2, alongside a roughly 30% cut in reasoning-token usage for agentic coding tasks — aimed squarely at token-efficiency for long coding sessions.

Simple version: Another big open-source coding model just got noticeably better and cheaper to run.

5. Google delays Gemini 3.5 Pro to July 17 after scrapping the 2.5 Pro architecture entirely

Rather than iterate on Gemini 2.5 Pro, DeepMind did a ground-up rebuild to fix gaps in math reasoning, SVG generation, and image quality. The new model is expected with a 2-million-token context window, a "Deep Think" reasoning layer, and pricing around $1.25/$10 per million input/output tokens — landing alongside Gemini 4 Flash and image model "Nano Banana Pro."

Simple version: Google's next big AI model got pushed back a few days because they rebuilt it from scratch instead of patching the old one.

6. Anthropic redeploys Claude Fable 5 after export controls are lifted

Fable 5 was pulled from global access on June 12 after researchers found a jailbreak bypass. Export controls lifted June 30, and Anthropic redeployed the model July 1 after training an improved safety classifier that blocks the exploit in "over 99% of cases," while partnering with Amazon, Microsoft, and Google on a shared jailbreak-severity framework.

Simple version: Claude's flagship model is back online after being pulled offline for weeks over a security issue.

7. ByteDance releases Seedream 5.0 Pro, a new image-generation flagship

ByteDance's latest image model claims to match or beat Western tools like Midjourney and Nano Banana on quality while undercutting them on price, leveraging TikTok's distribution to push adoption. It's pitched as a "reasoning" image model that understands design intent, not just prompts.

Simple version: TikTok's parent company has a new AI image generator that's cheaper and just as good as the top Western tools.

8. Boston Dynamics puts Gemini Robotics-ER 1.6 inside Spot

The integration gives Spot and the Orbit platform stronger spatial reasoning and autonomous decision-making for industrial inspection — including reading analog gauges at ~98% accuracy — without a human teleoperator, one of the clearest signs yet of frontier LLMs merging with commercial hardware.

Simple version: Google's AI now helps Boston Dynamics' robot dog inspect factories and read gauges on its own.

9. Anthropic in talks with Samsung for a custom Claude chip, prepping an IPO

Anthropic is reportedly discussing a custom AI chip tuned to Claude with Samsung, while preparing an S-1 filing for a possible IPO as early as October 2026. The company is on track for roughly $47 billion in annualized revenue and reportedly turned profitable this year.

Simple version: Anthropic wants its own custom chip and may go public on the stock market later this year.

10. Google caps Meta's access to Gemini models over compute constraints

Google reportedly limited how much Gemini capacity it will hand to Meta because it couldn't meet Meta's requested volume, delaying some of Meta's internal AI projects — a signal that raw compute, not model capability, is now the binding constraint at the frontier.

Simple version: Even Google doesn't have enough computing power to give Meta everything it's asking for.

11. Anthropic localizes Claude pricing for India, its second-biggest market

India accounts for 5.8% of global Claude usage. New INR pricing: Claude Pro ₹2,000/month (vs. $17 in the US), Claude Max ₹11,999/month (vs. $100), Team ₹2,399/seat. UPI payment support still isn't live, unlike ChatGPT's India offering.

Simple version: Claude subscriptions now cost rupees instead of dollars for Indian users — but you still can't pay with UPI.

12. PixVerse raises $439M, valuation tops $2B, pivots toward real-time interactive worlds

The video-generation startup's Series C extension, backed in part by Alibaba, funds a push beyond clip generation into real-time, interactive AI "world" and game-engine products — one of the largest AI video funding rounds this year.

Simple version: An AI video startup raised huge money to build interactive virtual worlds, not just short video clips.

13. Nous Research (Hermes) raising at a $1.5B valuation

The open-source agent maker behind Hermes — roughly 214,000 GitHub stars — is finalizing at least $75M led by Robot Ventures, with USV also participating, to expand its open agent tooling and hosted service.

Simple version: The team behind a popular free AI agent tool just raised a lot of money to grow it.

14. OpenAI reportedly pitched the US government a 5% equity stake in AI labs

Sam Altman floated a plan — reportedly worth $42.6B on OpenAI's own $852B valuation — for leading AI companies to allocate equity to a public fund modeled on the Alaska Permanent Fund, as OpenAI is said to be confidentially filing for an IPO amid its ongoing Apple trade-secrets lawsuit.

Simple version: OpenAI suggested giving the government a slice of ownership in AI companies to share the profits with the public.

15. TSMC posts record AI-driven revenue; Unitree Robotics clears IPO

TSMC reported Q2 revenue of ~$39.6B, up 36% year-over-year on AI chip demand. Separately, China's Unitree Robotics was approved for a $619M IPO on Shanghai's STAR Market to fund cheaper humanoid and quadruped robot development.

Simple version: The company that makes AI chips just had its best quarter ever, and a Chinese robot maker is about to go public.

Top 5 New / Popular AI Products

1. Viktor

AI Coworker

Topped Product Hunt's July 13 daily leaderboard with 598 upvotes — nearly double the runner-up — after raising $75M from Accel to build an AI agent positioned as a full "hire," not a tool, that proactively executes tasks inside Slack and Teams.

Why it's trending: It skips the "yet another dashboard" pattern — it just shows up in the channels teams already use and does the work.

2. AgentKey

Live Data Marketplace for Agents

Landed #2 on Product Hunt's July 13 leaderboard (322 upvotes), pitching itself as a one-stop marketplace where autonomous agents can buy and access live data feeds on demand.

Why it's trending: Solves a real pain point — agents need fresh, licensed data sources, and today that's a patchwork of APIs and scraping.

3. Osaurus

Local AI Agents for Mac

A Swift-built, fully offline macOS agent runtime hit #3 on Product Hunt (302 upvotes) and trended on GitHub, offering persistent memory and autonomous execution without sending data to the cloud.

Why it's trending: Growing appetite for local-first AI — no API bills, no data leaving the device, built for Apple Silicon.

4. Simba 3.2 (Speechify)

#1 Real-Time Voice Model

Speechify's Simba 3.2 took the top spot on Artificial Analysis's independent Speech Arena/TTS leaderboard, ranking above ElevenLabs, OpenAI, and Google DeepMind's voice models.

Why it's trending: An independently verified #1 benchmark result in a crowded voice-AI field is a rare, credible claim — not just marketing copy.

5. AI Media Buyer by Creatify

Autonomous Ad Management

Debuted on Product Hunt with 254 upvotes, offering autonomous ad-campaign management that Creatify says "gets smarter daily" by continuously optimizing creative and spend.

Why it's trending: Ad buying is tedious and metric-heavy — automating optimization loops appeals directly to small marketing teams and solo founders.