Two threads today: the industry's frantic hunt for cheaper inference — from OpenAI's secret 50% trick to Base44's own model — and the return of Fable 5 after 19 days offline. NLW closes with a practical guide to what you should actually try Fable on first.
Fable 5 is back — but the real story is what you do with it in the one free week.
After 19 days offline, Anthropic's Fable 5 returns for all global users, with 50% of weekly usage subsidized until July 7th. The conventional wisdom says save it for your hardest technical problems. NLW pushes back: from first-hand use, Fable is dramatically better at strategic thinking (it holds its ground under pushback instead of caving) and at rubric-driven writing (fewer AI-isms, better instruction following). Meanwhile the entire industry is racing to cut inference costs — the connective tissue behind OpenAI's secret technique, Base44's own model, and the shift toward open-weight deployment.
The Information reports OpenAI found an optimization technique that halved inference requirements for existing models, letting it serve all signed-out ChatGPT users on just 100 GPUs. The technique wasn't disclosed; speculation ranges from quantization to query routing — and none of those improve larger models without quality trade-offs.
NLW's caution: this is likely a smaller breakthrough than the headline suggests, and it's being tested on OpenAI's least-engaged users — which could be a reasonable first step or a hint of quality-degradation risk. Don't treat it as a silver bullet for the compute crunch.
— Stefanie Palazzolo, The Information. The Information's Stefanie Palazzolo argues OpenAI is holding the technique close because if it leaked, rival labs could quickly adopt it to lower their own costs.
— Everett Randall, Benchmark Ventures. Benchmark Ventures' Everett Randall's "AI mom test" captures why quality reductions may be tolerable for casual users — likely the exact segment OpenAI's new technique targets.
— Harry Stebbings, 20 Minute VC. 20 Minute VC's Harry Stebbings said founders from 10-person startups to a $200 billion public company all reported major inference savings in 24 hours: "The times they are a-changing."
Vibe coding platform Base44 launched Base1, fine-tuning an open-source base model on hundreds of millions of user interactions — the same playbook as Cursor's Composer. The bet: a model that only needs to be good at building web apps can rival the frontier while giving Base44 control over cost, latency, and quality.
— Shlomo, Base44. Base44's Shlomo frames the strategic logic: as AI becomes central to how software is created, controlling the model itself — paired with the harness — is a core moat.
AWS is investing a billion dollars in a new unit of forward-deployed engineers to help customers set up AI tools, joining OpenAI, Anthropic, Google, and Microsoft in the FTE race. It's upskilling salespeople into "solution architects" and focusing on healthcare, government, and financial services.
— Vasquez, AWS Frontier AI Engineering and Services. AWS's Vasquez noted the big enterprise shift toward AI budget optimization rather than just capability deployment, with open-source and open-weight models gaining ground.
Anthropic told Microsoft it plans to bring Claude Tag to Teams. It's an organization-centric agent — not tied to any user — with persistent memory and tool access, and it actually calls the full Claude Code suite rather than plain Claude.
— Satya Nadella, Microsoft CEO. Satya Nadella framed third-party agents like Claude as reinforcing Microsoft's ecosystem — though NLW wonders whether the tune changes as Claude gets more endemic across knowledge work.
To ease controversy over its Colossus data centers — powered off-grid by 46 gas turbines running without a permit — SpaceX is offering greater-Memphis residents half-price Starlink and free hardware, and recommitted to a wastewater treatment plant. The US government recently intervened to shut down the turbines, calling Colossus a national security matter.
NLW is glad data center operators are starting to cut communities into the benefits — but says making residents become customers to get a discount is far too small a gesture. Still, encouraging signs that these relationships are being rethought.
Anthropic announced the Department of Commerce lifted export controls, restoring Fable 5 for all global paid subscribers starting July 1st. A short subsidy extension covers up to 50% of weekly usage limits until July 7th; after that, access requires purchasing usage credits.
Anthropic said testing confirmed many less-capable models — including GPT-5.5 and Kimi K2.7 — could identify the same vulnerabilities Fable 5 flagged, and the reported technique only involved routine defensive cybersecurity work. It trained a new classifier claiming a 99% block rate, validated by Commerce's Center for AI Standards and Innovation as "extraordinarily strong."
— Dean Ball, policy advisor. Policy advisor Dean Ball welcomed Fable's return but noted no one knows what Anthropic actually did to make the model safe or how it applies to other models in the licensing queue. Still, he called a two-week review timeline "not insane" and real progress.
— Miles Brundage. Miles Brundage captured the prevailing policy mood: even with lots of open questions about what exactly Anthropic agreed to, people are cautiously optimistic just to have the model back.
NLW's read: there's a semblance of a practical framework now, but a lot of subjectivity goes into assessing risks. The best-case outcome is an efficient process with faster review for incremental updates — a bad outcome would be every capable release triggering the same weeks-long review and slowing the pace of breakthroughs.
The announcement's line that "some routine tasks like coding and debugging will fall back to Opus 4.8" sparked backlash that the relaunch was "fake." The Claude Code team clarified the wording was loose: only a small fraction of routine coding and debugging tasks get flagged and reverted to Opus.
Pitched as nearly as good as Opus 4.8 for a fraction of the cost, Sonnet 5 sits a few points shy of Opus on major coding benchmarks but posted a big jump on GDPVal, hinting at strong end-to-end agentic follow-through. NLW suspects the timing means Anthropic didn't expect Fable 5 back so soon.
Artificial Analysis found Sonnet 5 used ~40% more output tokens and ~3x the agentic turns of Sonnet 4.6; on the full benchmark it was more expensive than Opus 4.8 and even Fable. Without promotional pricing, it costs more per task than Opus — leaving many asking what it's actually for.
— Ben Davis. Ben Davis argued Sonnet 5 is good but inefficient and slow — spawning sub-agents, stacking PRs, reviewing itself adversarially. Others suggest the real framing is Fable 5 as the intelligent advisor and Sonnet 5 as the fast implementer running the sub-agents.
NLW's first-hand experience: unlike GPT-5.5 and Opus 4.8, which cave immediately under pushback and over-interpret instructions, Fable 5 accepts part of a critique while holding its ground on other parts — behavior he's never seen from another model. And it doesn't burn many tokens, so it won't chew through the 50% usage limit.
Contrary to Every's vibe check, NLW found Fable 5 much better at instruction-following, with fewer AI-isms and less try-hard prose. His suspicion: when you have a clear rubric or examples of good writing, Fable meets that standard far better than prior models — even if it's not necessarily better at blank-page writing.
Panjwani's advice: use Fable 5 for planning not implementation (delegate coding to GPT-5.5 via Codex), ask it for improvements on your most valuable projects, use other models to surface your hardest problems then have Fable propose solutions, and review Fable's output with GPT Pro. NLW agrees on the hard technical problems half but pushes hard on also using it for strategy and writing.