Fable 5 Goes Permanent, AWS Bills a Billion, and Open Source AI Closes the Gap

Good morning. It's Saturday, July 18, and we're covering Anthropic's reversal on Fable 5 subscription access, the AWS billing bug that showed some users trillion-dollar estimates, Mozilla's first State of Open Source AI report, and Moonshot's 2.8-trillion-parameter Kimi K3.

The Fable 5 saga ends with Anthropic keeping its flagship on Max plans. The AWS billing thread hit 1,193 points on Hacker News before the first fix shipped.

In today's brief:

• Fable 5 becomes permanent on Max and Team Premium starting July 20
• AWS Cost Explorer bug shows wildly inflated estimates - invoices unaffected
• Mozilla reports the open-closed model performance gap is now 3%
• Kimi K3 promises open weights for its 2.8T model by July 27

THE BIG ONE

Fable 5 Stays on Max Subscriptions

Anthropic announced on Twitter that Claude Fable 5 will be included in all Max and Team Premium plans starting July 20, at 50% of usage limits. Pro and Team Standard users keep access via usage credits and receive a one-time $100 credit.

The reversal ends weeks of uncertainty. Anthropic had originally planned to remove Fable 5 from subscription plans entirely after a series of deadline extensions - June 22, then July 7, then July 12, then July 19. The stated reason was compute capacity constraints, but competition from GPT-5.6 Sol and the just-released Kimi K3 made the math untenable: why pay $100 or $200 per month for a subscription that excludes your vendor's best model?

Simon Willison's analysis notes the capacity question remains open: "I wonder if they'll have to dial back their training efforts in order to make more GPUs available to help serve the model."

Why it matters: If you paused Fable 5 workflows or started migrating to alternatives, you can stop. The "Fablepocalypse" planning is over.

Our coverage: the June 22 decision checklist, how usage limits actually work, and Fable 5 vs Opus 4.8 for orchestration.

PLATFORMS

AWS Billing Bug Shows Users Trillion-Dollar Estimates

A bug in AWS Cost Explorer showed some users estimated bills of $1.7 billion, $2.5 trillion, or more - when their actual usage was under $5. The Hacker News thread hit 1,193 points and 708 comments as developers posted screenshots of absurd estimates.

The bug started around 7:38 PM PDT on July 16 and stemmed from "an issue with unit pricing within the estimated billing computation subsystem," according to Amazon's statement. Actual invoices were unaffected - this was the estimation layer, not the metering or billing system.

Amazon says corrected amounts should appear for all customers by noon Pacific on Saturday, July 18.

Why it matters: Estimated billing drives automation, alerts, and FinOps decisions. If you have scripts that trigger on billing anomalies, check whether they fired false positives this week.

RESEARCH

Mozilla: Open Source AI Performance Gap Is Now 3%

Mozilla released its inaugural State of Open Source AI report, built on analysis and a global survey of 950+ developers. The headline finding: the performance gap between open models and top proprietary systems like ChatGPT and Claude has narrowed to 3%, while inference costs have fallen up to 50x in three years.

The Hacker News discussion (446 points, 319 comments) focused on the report's revenue finding: open models now power roughly a third of real-world AI usage but capture only 4% of the revenue. The value is there, but it is not flowing back into the ecosystem that made it possible.

Other findings: MCP reached 97 million monthly SDK downloads and 10,000+ active servers in its first year, growing 4,750% in 16 months before being donated to the Linux Foundation's Agentic AI Foundation. The biggest gaps are now operational rather than capability-related - performance, integration, and maintenance tooling.

Why it matters: If you have been holding out on open models for quality reasons, this data suggests revisiting that assumption. The capability gap is largely closed; the remaining questions are operational.

Our coverage: the full Mozilla report analysis and model routers and the optionality advantage.

MODELS

Kimi K3: 2.8 Trillion Parameters, Open Weights by July 27

Chinese AI lab Moonshot AI released Kimi K3, a 2.8-trillion-parameter model they call the largest open-weight AI system ever built. It is available now via API at $3 per million input tokens and $15 per million output tokens, with open weights promised by July 27.

K3 uses a Mixture-of-Experts architecture with 896 experts, but only 16 are activated per token - roughly 1.8% of the pool. The model accepts text, image, and video input with a 1-million-token context window.

On Artificial Analysis benchmarks, K3 scores 57 on the Intelligence Index and ranks fourth out of 189 models - on par with Claude Opus 4.8 and GPT-5.5, behind only Claude Fable 5 and GPT-5.6 Sol. On the Frontend Code Arena, K3 took the top spot with 1,679 points, passing Fable 5 (1,631).

Simon Willison tested K3 with his pelican benchmark (a pelican SVG cost 25 cents at 16,658 output tokens) and noted the bigger limitation of that test: "The biggest limitation of the pelican is that it doesn't touch at all on the thing that matters most for today's models: agentic tool calling and the ability to operate tools reliably as conversations grow in length."

Why it matters: When open weights drop on July 27, teams running local or hybrid deployments get a Fable-class model to evaluate. The pricing is also notable - $3/$15 is the most expensive model from a Chinese lab, but still undercuts Western flagships.

Our coverage: Kimi K3 developer guide, K3 vs K2.7 comparison, and Kimi CLI vs Claude Code.

TOOLS WORTH A LOOK

1. Topcoat - Tokio's batteries-included Rust full-stack framework renders all markup on the server, supports Tailwind without Node, and translates expressions to both server Rust and client JavaScript. Early-stage, expect breaking changes. (OSS)

2. Open Book Touch - Open-source e-reader with a touchscreen, now on Crowd Supply. (hardware, preorder)

3. LLM cliche highlighter - Simon Willison's tool for flagging the ten most common patterns of LLM-generated writing. Useful for catching AI-written content before it ships. (free)

WHAT ELSE IS HAPPENING

• Kaiser nurses say AI surveillance is making patient care worse: 511 points on HN. Nurses report AI monitoring tools track keystrokes, bathroom breaks, and patient interaction times.

• TP-Link Kasa cameras leaked home GPS via UDP for 6 years: Unauthenticated UDP endpoint exposed exact GPS coordinates. 138 points. (security disclosure)

• The Zilog Z80 turns 50: The chip that powered the TRS-80, ZX Spectrum, and Game Boy is half a century old. 225 points.

• Julia Evans on running SQLite: Practical notes on WAL mode, busy timeouts, and the differences between SQLite and Postgres. 257 points.

• First atmosphere found on Earth-like planet in habitable zone: Astronomers detected an atmosphere on a rocky exoplanet in a star's habitable zone. 461 points.

• Recurse Center celebrates 15 years: The self-directed programming retreat launched via Hacker News in 2012. 607 points.

FROM THE SITE

What We Published Today

Our draft on securing AI coding agents is in progress - covering prompt injection through the toolchain, sandbox escapes like CVE-2026-50548 in Cursor, and the slopsquatting supply-chain risk. Defenses are ranked by effort: permission systems this week, OS-enforced sandboxes this month, and CI gates this quarter.

---

Every link above goes to a primary source or our sourced coverage. Sunday's brief lands when the news does - subscribe to get it by email.