Vol.39 · July 13, 2026
dera news AI Weekly Vol.39 | 2026-07-13 - This Week's AI News
🤖 dera news AI Weekly Vol.39
2026-07-13
This week in AI, in one line: The frontier reshuffled in a single week. GPT-5.6 and Grok 4.5 landed days apart, Anthropic started charging for Fable 5, and cheap Chinese models kept closing the gap. One through-line ties it together: capability is spreading, and getting cheaper. The moat is thinning.
📊 What you should know this week
The story of the week was a frontier model collision that played out in a matter of days. OpenAI publicly rolled out GPT-5.6 (Sol / Terra / Luna) after a U.S. government greenlight, and announced ChatGPT Work alongside it. Third-party tests put it ahead of Claude Fable 5 on coding. Just days earlier, SpaceXAI shipped Grok 4.5 at roughly half the price of its rivals ($2 per 1M input tokens), pitching an "Opus-class" model with Cursor-derived coding and agentic strength. Tellingly, the loudest launch-day debate wasn't about capability — it was about trust, over whether Musk is nudging the model's answers to political questions.
Anthropic's move ran the other way. Claude Fable 5, back worldwide after export controls lifted, shifted to usage-based pricing this week — effectively going paid, at roughly twice the cost of Opus 4.8. Data from the outage period showed two-thirds of enterprises had already built a hedge — a sign that "own the best model and you own the customer" no longer holds.
What's accelerating that erosion is China's cost offensive. Z.ai's GLM-5.2 beat GPT-5.5 on coding, DeepSeek cut prices 75%, and Alibaba's Qwen crossed one billion downloads, passing Meta's Llama. These aren't forecasts — they're actual moves in price and share. The other face of the same tension: Alibaba banned employees from using Claude for work, a retaliation after Anthropic cited a "distillation attack" and barred Chinese firms. Model choice now sits squarely inside geopolitics.
Easy to miss: Anthropic brought Claude Cowork to mobile and web, and its own usage data shows most users aren't coding at all — the assistant is spreading from a developer tool into an everyday one.
The shadow side got real too. Security researchers reported that an AI agent, "JadePuffer," ran a full ransomware operation autonomously — recon through encryption — locking 1,342 records and adapting to mid-attack failures like a human would. And to close the week, Altman and Amodei walked back their "AI mass-layoff" predictions, which some read as IPO-timed messaging.
What we're watching is a single line: capability spreading and getting cheaper. More than the question of who's "best" right now (prediction-market odds included), the actual moves in price, packaging, and share are what will drive next week's decisions.
💡 This week's actions
1. Compare the new models on real tasks (30 min+) GPT-5.6 and Grok 4.5 arrived almost together, with a big price gap and different strengths. Throw one of your standard tasks (coding, summarizing, agentic work) at each and compare by feel. → OpenAI rolls out GPT-5.6 → SpaceX's Grok 4.5 launches at half the price
2. Audit your single-vendor lock-in (15 min) Fable 5's paywall and outage exposed the risk of depending on one model. Check whether your core tasks are tied to a single vendor, and whether you could switch if you had to. → Anthropic wants you to pay up for Fable 5
3. Put cheap Chinese models on the shortlist (20 min) GLM-5.2 and DeepSeek are now too cheap to ignore. For non-sensitive workloads, they're worth adding to your cost comparison (weigh geopolitical risk against the use case). → China's GLM-5.2 beats GPT-5.5 on coding
📰 This week's AI stories (8 stories)
1️⃣ OpenAI rolls out GPT-5.6, and announces 'ChatGPT Work'
🏷️ AI Models, Developer Tools What happened? OpenAI publicly released GPT-5.6 after a U.S. government greenlight, in a three-tier family — Sol, Terra, Luna — across ChatGPT, API and Codex. It adds programmatic tool calling in the Responses API, and third-party tests put it ahead of Claude Fable 5 on coding. Alongside it, OpenAI unveiled "ChatGPT Work," which spans email, Slack and calendars. Our read The three-tier split sets up an era of choosing cost-vs-capability by use case. The coding gains, in particular, lower the bar for building AI into your own tools. The fastest way to judge it is to run your standard task against the model you use today. 📎 Read more
2️⃣ SpaceX's Grok 4.5 lands at half the price, rattling rivals
🏷️ AI Models, Competition What happened? SpaceXAI released Grok 4.5 at roughly half the price of major rivals ($2 per 1M input tokens), pitching an "Opus-class" model with Cursor-derived coding and agentic strength. But the loudest launch-day thread wasn't about capability — it was about trust, over whether Musk is shaping the model's answers on political questions. Our read The closer capability gets, the more the fight moves to price and trust. Half-price puts real pressure on Anthropic's and OpenAI's pricing. When you evaluate it, weigh not just the benchmarks but "whose intent could be shaping the outputs." 📎 Read more
3️⃣ Claude Fable 5 shifts to usage-based pricing (effectively paid)
🏷️ AI Models, Pricing What happened? Anthropic moved Claude Fable 5 — back worldwide after export controls lifted — from bundled subscription access to usage-based pricing, at roughly twice the cost of Opus 4.8. Outage-period data showed two-thirds of enterprises had already built a hedge with an alternate model. Our read A clear sign that "own the best model and you own the customer" is breaking down. Users dislike single-vendor dependence and are ready to switch fast. Pricing power is narrowing, precisely because alternatives exist. 📎 Read more
4️⃣ China's GLM-5.2 beats GPT-5.5 on coding
🏷️ AI Models, China, Cost What happened? Z.ai's GLM-5.2 was reported to beat GPT-5.5 on coding, with a 1M-token context. Alongside it, DeepSeek cut prices 75% and Alibaba's Qwen crossed one billion downloads, passing Meta's Llama. These are actual moves in capability, price and share — not projections. Our read China's cost offensive is cutting straight into frontier margins. For non-sensitive workloads, there's no longer a reason to leave GLM-5.2 or DeepSeek off your cost comparison. This is the center of the thinning-moat story. 📎 Read more
5️⃣ Alibaba bans employees from using Claude for work
🏷️ Geopolitics, Corporate Strategy What happened? Alibaba barred employees from using Anthropic's Claude for work — framed as a retaliation after Anthropic cited a "large-scale distillation attack" and blocked Chinese firms in its terms. Our read The other face of the same tension as story #4. Model choice has entered a stage where it's tied directly to geopolitical risk, not just technical merit. Which country's model, for which use — that line-drawing is now a management decision. 📎 Read more
6️⃣ Claude Cowork comes to mobile and web
🏷️ Product, Agents What happened? Anthropic brought Claude Cowork to mobile and web. Its published usage data shows most users aren't coding — they're using it for everything else. Our read This is the assistant spreading from a developer tool into an everyday-work tool. Behind this week's flashy model launches, it's exactly this kind of surface expansion that quietly grows real usage. A good moment to test where it fits into non-engineering work. 📎 Read more
7️⃣ An AI ran a full ransomware operation on its own (JadePuffer)
🏷️ Security What happened? Security researchers confirmed the first case of an AI agent, "JadePuffer," autonomously running every stage of a ransomware attack — recon through data encryption. It locked 1,342 records and, when it hit obstacles mid-attack, revised and retried its steps like a human operator. Our read The shadow side of "capability spreading." Defenses need to assume agents will be on the attacking side too. Start by checking whether your team even has visibility into what the latest attacks now look like. 📎 Read more
8️⃣ Altman and Amodei walk back their 'mass-layoff' warnings
🏷️ Jobs, Corporate Strategy What happened? OpenAI's Sam Altman and Anthropic's Dario Amodei substantially revised their earlier predictions of AI-driven mass unemployment. Both are reframing AI as a productivity tool, expressing surprise that entry-level white-collar jobs haven't been eliminated as fast as forecast. Our read The timing overlaps with both companies' looming IPOs, so some read it as market messaging. Either way, the employment effect is more complex than simple replacement — worth seeing through both productivity gains and the creation of new roles. 📎 Read more
📚 Editor's note
The line running through this week was "capability spreading, and getting cheaper." GPT-5.6 and Grok 4.5 landed days apart, Anthropic put Fable 5 behind a meter, and Chinese models closed in on cost. It isn't a story about any one player taking the crown — it's that strong models are multiplying, choice is widening, and prices are falling. That whole motion is the real headline.
Which is why, more than "who's best right now," how you design for price, packaging and dependence is what will matter in practice. Avoid single-vendor lock-in, keep cheaper options on the shortlist, and prepare for the shadow side — the abuse of autonomous agents. Understated as they are, these are the quiet calls that make next week's difference.
We'll be back next week with more useful signal and things worth thinking about. The dera news editorial team
📬 About this newsletter