Kimi K3 tops coding benchmarks; Fable joins Claude Pro plan

In a birthday editorial, Ben's Bites founder Ben Thompson reflected on a period of personal and professional transition—the arrival of a third child with health issues, a 7-month stint working at Factory, and seven months of reduced output. He described feeling disconnected from grand strategic pivots and instead recommitting to Ben's Bites itself, a newsletter he has sustained longer than any other project. Thompson resolved to inject more of his own voice, opinions on AI, and willingness to call out what he considers overheated claims, while testing new tools and sharing his learning process transparently.

Within that context, the newsletter highlighted Moonshot AI's Kimi K3 as a standout coding model. The 2.8-trillion-parameter model surpassed both Fable and GPT-5.6-Sol on Arena's frontend coding leaderboard and performed comparably on other benchmarks. Though impressive in capability, Kimi K3 carries a design tradeoff: it costs half as much as GPT-5.6-Sol per token but requires twice as many tokens, meaning overall costs balance out—Ben's Bites noted it is "an incredible model, but it's not an incredible value." The model supports images, features a 1M-token context window, and will have its weights released openly by 27 July. Currently available through Kimi's apps and API, new subscriptions are paused due to insufficient GPU capacity.

Anthropically moved Fable 5 into its core subscription lineup. Max and Team plan holders can allocate up to 50% of their weekly token limits to Fable. Pro subscribers ($20/mo) access Fable via usage credits and receive a one-time $100 credit as an incentive. Fable also demonstrated a mathematical breakthrough: it identified an example disproving the Jacobian conjecture, a 87-year-old open problem in mathematics. Separately, the newsletter noted that NotebookLM (Google's tool for extracting insights from documents) turned three years old and rebranded to Gemini Notebook, adding a Collections feature to organize multiple notebooks. Other tools gaining notice included Horizon (Sierra's agent for multi-week business outcomes charged by result rather than token), Cursor's rebuild of SQLite using multiple agent swarms (with cost comparisons showing Opus 4.8 + Composer 2.5 at $1,339 versus $10,565 for GPT-5.5 alone), and Notion AI's new browser capability in early access. The newsletter also showcased emerging patterns: Gumroad reported June 2026 as the first month it spent equal amounts on AI tokens and human payroll, with operations largely run by an AI called Gumclaw. Ben's Bites concluded by surfacing community discussion about different adoption levels—from people using AI only for search replacement to those running entire workflows through it—and noted a shift toward agents and AI coding as infrastructure rather than novelty.

The release of Kimi K3 and Fable 5's integration into Claude subscription tiers reflects a maturing market for AI coding assistance. Kimi K3's strong performance on Arena's coding benchmarks—surpassing both Fable and GPT-5.6-Sol on frontend tasks—demonstrates that competition among coding models is intensifying, with Beijing-based Moonshot AI now offering a credible alternative to Western incumbents. The tradeoff between Kimi K3's lower per-token cost and higher token consumption (half the price per token but double the tokens used) shows that efficiency gains alone do not guarantee value—a theme Ben's Bites editorial staff flagged directly.

Anthropically's decision to embed Fable 5 permanently into Max, Team, and Pro tiers signals confidence in the model's maturity and a shift in business model logic: rather than treating AI coding as a premium add-on, it is becoming a baseline feature. The $100 one-time credit for Pro users and the 50%-of-limits cap suggest Anthropic is managing usage incentives to balance accessibility with cost. Kimi K3's forthcoming open-source release by 27 July will likely accelerate adoption among developers who prefer local or self-hosted infrastructure, particularly outside regions where cloud API costs are prohibitive.

No comments yet. Be the first to share your thoughts!