01AI$Fable 5's free ride ends Sunday. Claude's top model drops out of subscription plans July 19 and moves to $10/$50 per million tokens. Third deadline extension in five weeks.
02$OpenAI made the same move nine days earlier, more quietly. Excel and Sheets agents moved to token-based credit pricing on July 6. If your finance team builds models in ChatGPT, the bill just got variable.
03AIGoogle missed its own deadline again. Gemini 3.5 Pro was rebuilt from scratch after failing internal reliability tests and now targets a stopgap launch today, its third delay.
04AIA frontier lab shipped a model designed to lose. Mira Murati's Thinking Machines released Inkling, its first model, and says outright it isn't the strongest one available.
05AIWashington's AI rulebook is splitting in two. Anthropic is underwriting state safety laws while OpenAI fights for one federal standard, both in front of IPO investors.
06$DeepSeek's API clock is ticking. Anyone with deepseek-chat or deepseek-reasoner wired into production has until July 24 to migrate before the old endpoints stop working.
LEAD STORY
The Fable 5 Cliff
Anthropic has now moved the goalposts on Claude's flagship model three times in five weeks. Sunday is when the bill actually comes due.
Anthropic's most capable model has been free to use for anyone on a paid Claude plan since it returned to general availability on July 1. That arrangement ends Sunday, July 19, at 11:59 PM Pacific. It is the third time Anthropic has pushed the deadline back, first from July 7 to July 12, then from July 12 to July 19, each time citing capacity rather than demand. After Sunday, Fable 5 usage beyond a subscription's normal weekly pool runs on prepaid credits billed at $10 per million input tokens and $50 per million output tokens.
Fable 5 Deadline, Extended Twice
Confirmed dates from Anthropic's own support documentation
The mechanics matter more than the headline. Fable 5 draws down a subscriber's weekly usage pool faster than Sonnet 5 or Opus, so the free window was never really free, it was a faster-draining allowance. Once that allowance is exhausted, or once Sunday passes, the choice is pay per token or fall back to a cheaper model. For any workflow currently routed to Fable 5 by default, that is a cost decision that needs to be made this weekend, not discovered on next month's invoice.
"Fable 5 won't permanently leave subscriptions, Anthropic says, once it has enough compute." Read: this is a supply problem, not a pricing experiment, and it's the second time in three weeks a frontier vendor has quietly repriced its most-used product.
What Fable 5 Costs After Sunday, Against the Field
Dollars per million tokens, input vs. output
Input $/M tokensOutput $/M tokens
Fable 5 usage-credit pricing takes effect July 20. GPT-5.6 figures are OpenAI's three published tiers. Sources: PYMNTSAnthropic support doc
The read for operators: if a workflow currently leans on Fable 5 for anything high-volume, routing stops being a nice-to-have this weekend and becomes a line item. See The six50 POV below for what we'd do about it.
FINANCE DESK
The Flat-Rate Era Just Ended, Twice
Two of the largest AI vendors moved off flat pricing for their most work-heavy products in the same two-week window. If your finance team runs on either one, this is the week to notice.
Nine days before Anthropic's Fable 5 deadline, OpenAI made a quieter but arguably bigger move. On July 6, OpenAI's Enterprise and Edu release notes confirmed that Workspace agents and Excel and Sheets tasks are shifting to token-based credit pricing. Cost now scales with input tokens, cached input tokens, and output tokens, meaning with what the agent actually does, not with whether it was invoked at all. For any SMB finance team using ChatGPT to build models, reconcile data, or draft a board deck in Excel, the monthly bill just became a function of how hard the agent worked, not how many seats were purchased.
✓ Still Flat-Rate
Standard ChatGPT and Claude chat, within plan limits
This is not a coincidence of timing, it is the same structural shift showing up at two different vendors within two weeks of each other. Per-seat pricing rewarded stable headcount. Agentic workflows reward, and now charge for, variable consumption: a finance team that automates month-end close will burn dramatically more tokens than one that just chats with a model occasionally. Both vendors are converging on charging for the work performed, not the access granted.
six50's 4-Point AI Cost Control Checklist for SMB Finance Teams
1
Set a monthly token or credit ceiling per workflow, not per department. A ceiling on "AI" as a category is too blunt to catch a single runaway automation.
2
Route routine tasks (drafting, first-pass review, summarizing) to the cheapest model that clears the bar. Save premium tiers for judgment calls that actually need them.
3
Track cost per completed workflow, not adoption or run count. That is the same discipline you'd apply to any other line-item vendor spend.
4
Revisit model routing quarterly. Pricing structures are moving fast enough that a routing decision from Q1 may already be stale by Q3.
The read for operators: this is exactly the blind spot the First 90 Days Diagnostic's AI automation roadmap is built to catch. A close process that looks automated and cheap on the surface can be one pricing update away from a very different number, and most SMB operators won't notice until the invoice lands.
COMPETITIVE LANDSCAPE
Google Misses Its Own Deadline, Again
Gemini 3.5 Pro was supposed to ship in June. Then July. The rebuild that followed tells you more than the launch will.
Google DeepMind scrapped Gemini 3.5 Pro's original architecture after Vertex AI enterprise testing turned up structural failures, not polish problems, in recursive tool-calling, SVG generation, and mathematical reasoning. Rather than patch the existing model through post-training, the team rebuilt it from the ground up. Sundar Pichai told developers at I/O in May to "give us until next month." That month came and went. A stopgap release is now targeted for today, July 17, the third missed internal deadline, with pricing and the full model card still unconfirmed as of this writing.
3Missed Deadlines
2MToken Context (Reported)
~$250/mo Ultra Tier (Est.)
SWE-bench Pro: Where the Field Actually Stands
Coding benchmark, percent resolved, as of July 16 leaderboard
Gemini 3.5 Pro omitted, no verified score published as of this issue. OpenAI's own July audit flagged roughly 30% of the public SWE-bench Pro task set as broken and withdrew its earlier endorsement of the benchmark, treat these as directional. Source: BenchLM.ai, Jul 16 2026
The read for operators: the benchmark race is noisy and self-interested on all sides. What is not noisy is that Google, the deepest-pocketed lab in this fight, still couldn't ship on time twice in a row. Treat any lab's release date the way you'd treat a contractor's timeline: pad it, and don't rebuild a workflow around a model that hasn't shipped yet.
MODEL RELEASE
The Model Designed to Lose the Leaderboard
Thinking Machines had one shot at a first impression. It spent it on a model built to be reshaped, not admired.
Mira Murati's Thinking Machines released its first model this week: Inkling, a 975-billion-parameter mixture-of-experts system that only activates about 41 billion parameters per task. It reasons natively across text, image, audio, and video, carries a 1-million-token context window, and lets users dial reasoning effort up or down. It is also, per the company's own announcement, open-weight, meaning any enterprise can download and modify it directly, something none of the closed frontier labs offer.
975B Total, 41B Active
Inkling's mixture-of-experts split, in billions of parameters
What makes Inkling worth a second look is what the company chose not to claim. Thinking Machines states plainly that Inkling "is not the strongest overall model available today, open or closed." It is positioning the model as a starting point for fine-tuning through Tinker, its customization platform, rather than a ChatGPT or Claude competitor.
A base model that's explicitly average out of the box, built for you to specialize, is a different product category than a frontier chatbot. Priced and positioned right, it's a category SMB operators may actually prefer.
The read for operators: watch what gets built on top of Inkling over the next 60 days, not its day-one benchmarks. A customizable open-weight base that a business can shape to its own workflow can beat a closed frontier model on cost and control, even while losing every leaderboard.
WORTH KNOWING
POLICY
Anthropic and OpenAI are fighting the AI law battle from opposite ends
Anthropic keeps backing tougher state AI safety bills (California's SB 53, New York's RAISE Act, an Illinois audit requirement) while OpenAI pushes for one federal standard to preempt all of them. Both companies are doing this in front of investors ahead of expected Q4 IPOs, which makes regulatory posture a disclosure question, not just a policy stance.
OpenAI issued its first public response to Apple's allegations, saying it has seen no evidence supporting the trade secret theft claims. No further detail released.
deepseek-chat moves to deepseek-v4-pro and deepseek-reasoner moves to deepseek-v4-flash on July 24. Anyone with DeepSeek calls hardcoded into a pipeline should be auditing that now, not on the 24th.
What we'd tell a $2M-$50M operator to do with today's news
01 — MODEL ROUTING
Audit what's running on autopilot before Sunday. If Fable 5 or another top-tier model is the default for routine drafting or summarizing, that's the same mistake as running every invoice through your most expensive software seat. It's the first thing we check in a First 90 Days Diagnostic.
02 — COST GOVERNANCE
Put a dollar ceiling on agentic workflows, not just a seat count. OpenAI's July 6 shift to token-based Excel and Sheets pricing means your close process now carries a variable cost. Cap it by workflow before finance finds the number on next month's invoice, using the checklist above.
03 — VENDOR SELECTION
Stop assuming the frontier model is the right model. Inkling's own pitch, fine-tune us instead of trusting our benchmark score, is the same logic behind our AI automation roadmaps. Most SMB workflows need something smaller, cheaper, and controllable, not the leaderboard leader.
04 — REGULATORY WATCH
Don't overreact to the Washington fight yet. Anthropic versus OpenAI on state AI law is an IPO story before it's an SMB compliance story, unless you're already in a regulated vertical where a vendor's posture affects your own audits.