Fablepocalypse Cancelled, Tibo Does It Again & the Real Price of Kimi K3

Claude Code & Anthropic Updates

Fablepocalypse cancelled: Fable joins Max and Team plans for good

The week-long "use your Fable allowance before it vanishes" saga ended with Anthropic announcing that beginning July 20, Claude Fable 5 is included in all Max and Team Premium plans at 50% of limits, with Pro and Team Standard users keeping access via usage credits plus a one-time $100 credit. Anthropic's Thariq (@trq212, 282K views) framed it as "a heroic effort by many people at Anthropic working sometimes literally around the clock" — it wasn't clear they'd secure capacity in time to avoid a gap in Max access. The replies are a great temperature read: plenty of gratitude, but the top reply thanks "GPT 5.6 Sol and my boy Kimi K3 for the extra fable limits," and several users vent about the FOMO-driven comms — people burned through their allowance racing a deadline that then evaporated. A biologist in the replies also notes Fable's safety classifiers block their legitimate scientific work, pushing them to Codex.

Simon Willison's take was more relaxed: "Huge relief, the Fablepocalypse has been permanently cancelled". Jerry Liu, as a Max user: "you know what, i'll take it". Armin Ronacher's counterpoint: "Slacking off with fable is expensive."

Theo: "FABLE IS THE SAME MODEL AS MYTHOS"

Tired of influencer misinformation, Theo posted an all-caps explainer (209K views): Fable 5 and Mythos 5 are the same weights on the same servers — "two separate entrances to the same place." Fable's extra safeguards are layers in front of the model (request checks plus response monitoring), while Mythos is the low-friction entrance for trusted parties. Best analogy from the replies: "TSA-Fable vs TSA Precheck-Mythos, lead to the same terminal if you get through." Theo also flatly shut down the recurring conspiracy that the current Fable is a different model from the one available at launch: "It is."

Simon Willison begs for a Claude Code on the web regression suite

Simon publicly asked Anthropic for "an automated test suite somewhere that ensures Claude Code on the web never blocks me from cloning or interacting with other public repos from within my existing sessions" — his "Clone repo X from GitHub to /tmp for reference" pattern is one of his most-used prompts, and this isn't the first regression. The replies are substantive: one user notes cross-repo pulls are missing from Cursor, Codex, and Claude cloud agents ("makes me want to create a mega monorepo"), another dissects the screenshot into two separate failures (proxy choking on release binaries vs. the sandbox refusing a second repo owner), and several echo the deeper point — guardrails tuned by policy instead of context will always over-block, and "false-positive evals matter as much as capability evals."

Kimi K3 Reality Check & Model Wars

Theo's value math: K3 costs the same as GPT-5.6 Sol, at ¼ the speed

Day two of Kimi K3 brought the correction phase. Theo's most-discussed post of the day (325K views, 230 replies): "Kimi k3 is an incredible model. It is not an incredible value." K3 is half the price of GPT-5.6 Sol per token, but 5.6 uses half as many tokens — so price evens out — and 5.6 runs 2x the TPS, getting work done ~4x faster at the same cost. He still loves K3 ("killer for a bunch of fun stuff like using Blender and making dumb games"), he's just tired of people pretending it's cheaper. Notably he says K3 is genuinely token-efficient — fewer tokens than any Anthropic model, with Sonnet 5 being "10x more token hungry than it should be" — and warns the upcoming open-weights release (July 27) will only shave 10–20% off hosting costs. Victor Taelin agrees in the replies: smarter than GPT-5.6 "in most ways that matter to me," but too slow to be practical.

The $19 Kimi Code plan: one prompt, quota gone

LLMJunky's hands-on backs this up: the rate limits on Kimi Code's $19 plan are "really bad" — one prompt building a browser game hit the 5-hour quota at ~85% complete and burned 20% of the weekly allowance ("Fable one shot this for me on the $20 plan"). His Rocket League benchmark verdict: K3's version was "pretty mid," 4th or 5th best he's seen, behind Opus 4.8 — nice UIs, disappointing physics. And Armin Ronacher adds a caveat for evaluators: K3's current max-thinking-mode restrictions make it a bad match for some basic tasks.

But K3 tops the Frontend Arena — and Jerry Liu sees distillation heaven

Meanwhile K3 hit #1 in the Frontend Code Arena, surpassing Claude Fable 5 with a 17-place jump from K2.6, ranking first in 6 of 7 domains (only Gaming went to Fable). Jerry Liu's spicy framing: the proliferation of Chinese open-source models winning task-specific benchmarks "may not be great for American AI dominance," but it's fantastic for everyone else — "Every American AI startup will survive by distilling from the latest open-weight models, instead of solely depending on the frontier labs."

Steipete: don't trust benchmarks, Terra high beats Sol low for reviews

Peter Steinberger switched his GitHub review bot @clawsweeper to GPT-5.6 Terra high: ~40% faster than 5.5 with negligible quality loss and massively cheaper (xhigh negated the perf wins without moving review evals). Filed under "don't trust benchmarks": for his issue/code-review use case, Terra high delivers far better results than Sol low.

Codex & OpenAI

"Oops... I did it again" — Tibo resets everyone's limits

OpenAI's Tibo (@thsottiaux) reset usage limits for all paid Codex and ChatGPT Work users (873K views, 14.6K likes), quoting his own "GPT-5.6 Sol confirmed to be an extremely good model". He noted the reset transitively "might also have reset other rate limits out there" — a wink at Anthropic's same-day Fable announcement. The reply section is a case study in reset-economics: multiple users lament having spent their banked resets hours before the global one dropped, and there are feature requests for banked resets that auto-apply at expiry. The timing war with Anthropic's Fable news was lost on nobody: "right after they announced Fable will stay in plans 😂"

Simon Willison connected a different dot: "ChatGPT Work" is really "OpenAI Claw".

Codex corner

Agentic Coding & Workflows

Other Notes