2026-07-14

1. Anthropic Honeycomb Spotted in Cursor — Unreleased Model, 500K Context, March 2027 Cutoff

On July 8, a Cursor user discovered "claude-honeycomb-20260707" accessible via a shared Anthropic account integration. The model has a 500K token context window (2.5× Fable 5's 200K), a training cutoff of March 2027 (implying it was trained on data through Q1 2027 — post-Fable 5), and benchmarks placing it above Fable 5 on early internal evals. Community interpretation: Honeycomb is either Fable 5.5 (tuned), a specialized variant of the Mythos stack, or an early Sonnet 6 precursor. Anthropic has not commented. The model did not appear in any Anthropic changelog or announcement.

Why it matters: The 500K context would be the largest publicly detectable Anthropic context window, beating Sonnet 5's 1M (API-gated) and exceeding what's visible in Claude.ai. If Honeycomb is Mythos-class, it suggests Anthropic is quietly testing a second Mythos deployment before any announcement. The March 2027 knowledge cutoff also implies faster iteration cycles than the public-facing roadmap suggests.

→ Anthropic · Source: sources/x/2026-07-08-anthropic-honeycomb-eap-cursor.md

2. Google DeepMind Nano Banana 2 Lite — 3B On-Device Multimodal, 160 Languages, 10× Efficiency

Released June 30, 2026 — missed in 14 consecutive ingests. Nano Banana 2 Lite is Google DeepMind's new 3B on-device multimodal model (text + image), part of the Gemma 4 family. Key stats: 160-language support, 10× efficiency improvement vs Nano Banana 1, INT4 quantized to ≤3GB RAM (fits budget Android hardware), Apache 2.0 license. Targeted at mobile inference on devices with no cloud dependency. Fills the gap between Gemma 3n (5B, Apr 2026) and fully on-device multimodal for emerging markets.

Why it matters: 160-language support is the broadest in any sub-5B open-weight model. The 10× efficiency jump means Nano 2 Lite runs on phones where Nano 1 would not. This is Google's play for the 6B+ users outside the premium smartphone market — far outside OpenAI's or Anthropic's current reach on-device. Emerging-market AI access is a compounding advantage; each new voice/language adds usage data.

→ Google DeepMind · Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) · Source: sources/blogs/google-2026-06-30-nano-banana-2-lite.md

3. Fable 5 Third Extension to July 19 — Anthropic Announces 90 Minutes Before Deadline

On July 13, just 90 minutes before the July 12 "credits-only from July 13" deadline, Anthropic announced a third extension of free Fable 5 access through July 19, 2026 at 11:59 PM PT. Terms: Pro/Max/Team/Enterprise subscribers receive a +50% rate-limit boost at no extra cost. After July 19: credits-only at $10/$50 per Mtok. Anthropic cited "continued capacity expansion work." This is the third consecutive extension since the June 30 export-control lifting (June 30 → July 7 → July 12 → July 19).

Why it matters: Three extensions in 19 days signals that Anthropic is either struggling to stand up capacity (Colossus 1 lease to Anthropic begins Q3 — timing may not be coincidental) or is deliberately staging the credits-only transition to maximize adoption stickiness. The last-minute July 13 announcement suggests the deadline decision is being made at the wire each time. The +50% rate limit boost is notable: it suggests actual free capacity is available but is being rationed. July 19 is almost certainly the real cutoff unless Colossus 1 GPU allocation is delayed.

→ Anthropic · Claude Fable 5 · Source: sources/blogs/anthropic-2026-07-13-fable5-july19-extension.md

4. OpenAI Removes ChatGPT Work 5-Hour Cap — Integrations + Memory Now Unlimited

OpenAI removed the 5-hour daily usage cap for ChatGPT Work ($30/month) on July 13. Simultaneous additions: built-in Slack, email, and calendar integrations (no connector setup), persistent memory across Work sessions, and background task execution that runs after the user closes the tab. Previously, Work was the only tier with a hard time cap — its removal aligns Work more directly with the Enterprise tier for power users who don't need admin/SSO features.

Why it matters: The cap removal plus background tasks turns ChatGPT Work into a persistent work agent, not just a chat window. Combined with the Slack/calendar/email integrations, it's targeting exactly the kind of "ambient AI assistant" that Google Gemini Spark was designed for. At $30/month vs. Gemini Spark's Google Workspace bundling, the pricing fight here is real. This also validates the move toward unlimited agentic tiers — the cap was the main differentiator forcing enterprise contracts.

→ OpenAI · Source: sources/blogs/openai-2026-07-13-chatgpt-work-cap-removal.md