AI Pulse

Sol vs. Fable 5, ChatGPT Work vs. Cowork, and Zendesk's Quiet July Upgrade

By Felix Maru · July 14, 2026 · 6 min read

In the span of four days, the two biggest AI labs traded blows and both rounds landed on the same place: your support stack's cost model and your agents' daily workflow. Fable 5 goes from free to pay-per-token on July 20. OpenAI shipped Sol and a new agentic work product. And Zendesk dropped a July update that most support teams haven't read yet, but should. Here is what actually happened and what your team needs to do about it.

Fable 5 Is Free Until July 19. Then the Math Changes.

On July 13, Anthropic extended free access to Claude Fable 5 for subscribers through July 19, the second extension in a single week. The reason, as reported by Forbes on July 13: OpenAI launched GPT-5.6 (Sol) on July 9 and Anthropic responded by resetting weekly limits and handing subscribers another seven days of its most capable model at no extra cost.

After July 19, Fable 5 usage moves to prepaid credits: $10 per million input tokens and $50 per million output tokens, confirmed by Dataconomy. At that rate, a team running a drafting flow across 10,000 tickets a month is looking at roughly $300 per month for a single automation lane. That multiplies across triage, summarization, and any other flows hitting the same model. The shortcut of defaulting to Fable 5 everywhere because it was free disappears on July 20.

Three things worth doing before then: map every automation calling Fable 5 and estimate the monthly token cost at new pricing. Check whether each call actually needs Fable's depth, or whether a lighter model would produce the same output. Then decide: keep Fable for the flows where its reasoning genuinely matters, and move simpler classification tasks to a cheaper model. That decision is easier to make calmly this week than reactively after the first invoice arrives.

Sol Is OpenAI's Answer. Here's What It Actually Does for Support.

OpenAI released GPT-5.6 on July 9, with three tiers: Sol (flagship), Terra (mid), and Luna (base). OpenAI describes Sol as its reasoning model for "complex work across coding, research, science, cybersecurity, and design," per the OpenAI help documentation. Anthropic's July 13 extension of Fable 5 was a direct response to Sol's launch, per both Forbes and Neowin.

For support teams, the benchmark race between these two labs is less interesting than the practical question: what does each model actually do on your real ticket mix? Sol's strength is in structured reasoning and multi-step analysis, which is useful for complex escalation summaries, troubleshooting decision trees, and cross-system data lookups. Fable 5 has shown genuine strength in tone calibration for empathy-heavy conversations and brand-consistent draft replies, the kinds of outputs that go directly in front of customers.

These are different strengths, not a clean winner. If your team is evaluating or considering switching, this week is actually a good window for a structured comparison. Pull 20 to 30 representative tickets from your queue across different types, run them through both models, and have agents score the outputs on resolution accuracy, tone appropriateness, and how much editing was required. One afternoon of real testing tells you more than any benchmark chart or press release.

ChatGPT Work and Cowork: The Agent Race Goes Async

On July 7, Anthropic announced that Claude Cowork, previously desktop-only, now runs on the web at claude.ai and on iOS and Android in beta, starting with Max plan subscribers, per TechCrunch on July 7 and 9to5Mac on July 13. The key technical shift: remote sessions run on Anthropic's servers, so work continues with no device online. Start a task from your desk, get a status notification on your phone, and pick up the finished output later even if your laptop is closed.

OpenAI shipped a direct counter with ChatGPT Work alongside GPT-5.6. The OpenAI product page describes it as an agent that can "research and analyze information, work across connected apps and files, and create finished documents, spreadsheets, presentations, reports, and Sites."

For support operations, both tools point at the same practical use case: AI agents completing structured work asynchronously, without someone actively managing a browser tab. Think overnight generation of weekly ticket-volume summaries, end-of-shift escalation reports, or knowledge-base draft articles queued while the team is offline. Removing the "must keep a device running" constraint is the real unlock. Plenty of support teams have been cautious about async AI workflows precisely because they felt like they needed to supervise the run. That friction is gone now.

The human review checkpoint is not optional. It is the point. Any agent completing a draft or summary still needs a person to read and approve it before it reaches a customer or drives a decision. Build that checkpoint in deliberately from the start: who reviews, at what point, and what triggers escalation. That structure separates a useful async workflow from one you turn off three months later because something slipped through.

Zendesk's July Pack: The Operational Upgrade Most Teams Missed

While the model war dominated the conversation, Zendesk's July 2026 release notes included several features with more immediate day-to-day value for most support teams than anything in the AI lab press releases.

AI-based predictive routing is now live for messaging tickets in omnichannel routing. Instead of assigning by round-robin or first-available, the system predicts which agent is most likely to resolve a ticket fastest given their current workload and the ticket's context. Zendesk also added Agent Engagement Time as a new metric: the total time an agent actively works a ticket from assignment to close, giving teams a meaningful way to measure whether predictive routing is actually improving throughput.

Why this matters more than people give it credit for: most queue backlog is not purely a staffing problem. It is often a matching problem. A billing dispute routed to an agent who has handled fifty similar cases this week resolves faster and with better quality than the same ticket sent to whoever happens to be free. Smarter routing raises your first-contact resolution rate without adding headcount.

Forethought AI agents are now available as a purchasable add-on inside Zendesk, covering autonomous inquiry resolution, intent identification, task automation, and agent assistance directly within the platform. The native Contact Center is now generally available, integrated into the Agent Workspace with single status management, call monitoring, and expanded language support. And Workforce Management is now included in Suite Professional at no extra cost, with 24-hour insights into agent activity and average handle time.

None of these are splashy announcements, but together they move Zendesk closer to a native AI workforce platform. If you are on Zendesk today, check your plan entitlements before renewing or purchasing a separate routing, forecasting, or WFM tool. You may already have what you were about to pay extra for.

What to Actually Do This Week

Three things, none of them abstract:

The AI model war moves fast. The best response to that speed is not to track every announcement or declare a winner. It is to know exactly what your support desk requires, test against that baseline with real tickets, and make decisions from your own data. The model that wins a benchmark may not be the one worth paying for at your ticket volume, ticket type, and quality bar. Only your queue can answer that.

Sources

Thinking through which AI model to commit to for your support automation, or trying to make sense of what your Zendesk plan now covers? Drop me a line and I will share what I have seen work at different ticket volumes and types.

Share 𝕏 in ✉

Comments