Price:

AI & TECH

Anthropic nerfs AI research queries while Grok becomes a developer tool

Friday, July 17, 2026 · from 3 podcasts, 4 episodes
AI Daily BriefNerd SnipeMoonshots with Peter Diam…
  • Grok 4.5 reliably completes complex coding tasks, creating a third option for developers.
  • OpenAI shuttered Codex, folding developer tools into a bloated ChatGPT super app.
  • Anthropic throttles biology and AI research queries on Fable 5, angering open-source researchers.

Anthropic’s Claude Fable 5, released July 9th, is the most capable AI model to date. Nathaniel Whittemore cites its performance across coding benchmarks: 80.3% on Swebench Pro versus GPT-5.5’s 58.6%, and a 29.3% score on Frontier Code versus Opus 48’s 13.4%. This gap moves AI from task execution to managing autonomous, long-running responsibilities.

Yet Anthropic nerfed the model’s ability to assist with frontier LLM development, blocking infrastructure queries about pre-training pipelines. Researchers Ellie Bau, Nathan Lambert, Dean Ball, and Will Brown called the invisible throttling misaligned and hostile to open model research.

"Fable 5 initiates a third AI era: moving from asking questions to assigning tasks, and now to giving responsibilities."

- Felix Ryberg, The AI Daily Brief

The same week saw the collapse of the OpenAI-Anthropic duopoly. Alex Gleas argues Grok 4.5, GPT-5.6, and Muse Spark pushed four American labs onto the optimal performance frontier. Grok lacks frontier intelligence but excels at reliability. Theo and Ben argue it handles multi-part coding tasks without hallucination loops, offering a viable third option for production workflows.

OpenAI pivoted away from developers. Ben describes OpenAI folding the standalone Codex app into a unified ChatGPT desktop experience. Theo calls it “slop design,” a crowded dashboard that obscures tools developers once trusted.

Apple sued OpenAI July 13th over hardware IP theft. The 41-page complaint alleges former Apple hardware chief Tan Tan retained access to confidential files after poaching 40 employees. This legal front marks Big Tech and frontier labs as direct competitors for the post-smartphone device market.

The real stake isn’t model intelligence but economic control. Anthropic moved Fable 5 to pay-per-use pricing June 23rd, ending the unlimited subscription era. Microsoft’s Mustafa Suleyman pitches frontier tuning as an alternative to mass consumption, aiming to keep agentic costs from crushing enterprise margins.

"Users are moving from giving tasks to giving responsibilities. Instead of asking the AI to fix a bug, they set it on a loop to autonomously monitor crashes."

- Felix Ryberg, The AI Daily Brief

Friction defines the new market. Developers trade frontier smarts for Grok’s consistent $2-per-million token price. Researchers trade Anthropic’s top-tier model for its safety dragnets. OpenAI trades its clean Codex brand for a cluttered super app. The competition is about what you’re willing to lose.

#Models#Coding#Enterprise
Nathaniel WhittemoreNathan LambertMustafa SuleymanAnthropicOpenAIClaudeGrokCodexChatGPTSparkOpusGPT-5Microsoft

Source Intelligence

- Deep dive into what was said in the episodes

The AI Daily Brief: Artificial Intelligence News and Analysis
The AI Daily Brief: Artificial Intelligence News and Analysis

Nathaniel Whittemore

5 AI Engineering Trends for Non-Engineers • Jul 15

  • Anthropic launched Claude Fable 5, its first Mythos-class model, positioning it above Opus. Mythos 5 lacks Fable’s controversial safeguards and is initially only available to Project Glasswing partners.
  • Nathaniel Whittemore calls Fable 5 the best AI model ever available but notes exploiting state-of-the-art models now requires more than simple prompts.
  • Fable 5 dominates benchmarks: 78% on ExploitBench versus GPT55’s 34%, 66% on HealthBench versus GPT55’s 51.8%, 13.3% on the legal agent benchmark versus GPT55’s 2.1%, and 1932 on GDP Val’s knowledge work test versus Opus 48’s 1890.
  • The model excels at agentic coding: 80.3% on Swebench Pro versus GPT55’s 58.6%, 88% on Terminal Bench versus GPT55’s 83.4%, and 29.3% on Frontier Code versus Opus 48’s 13.4%. Fable scored 91% on Every’s Senior Engineer benchmark.
  • Artificial Analysis found Fable 5 topped its blended benchmark run, overtaking Opus 48 and GPT55, though some noted the overall gap was only five points.
  • Fable 5 scored 72.9% on Cursor Bench, eight points above the previous best, but is more expensive on that cost-performance test.
  • Cognition’s Frontier Code benchmark aims to assess real-world coding quality for merging into production, not just passing unit tests. Meter found more than half of Swebench results are unmergeable 'slop'.
  • Users report Fable flags basic biology terms like 'mitochondria' and 'cancer' as biosecurity risks, switching to Opus 48. Creo, Daria Anupmas, and Fernando documented these blocks.
  • Anthropic’s system card reveals new interventions limiting Fable’s effectiveness for frontier LLM development tasks like building pre-training pipelines or ML accelerator design, aiming to prevent aiding competitors.
  • Researchers Ellie Bau, Nathan Lambert, Dean Ball, and Will Brown criticize the invisible nerfing of AI research capabilities, calling it sad, misaligned, hostile, and a barrier to open model research.
  • Anthropic mandates a 30-day data retention and review policy for Mythos-class models on all platforms. Mike Taylor warns this violates NDAs if memory is on, pulling historical chats into review.
  • Users debate token efficiency: Theo and Chubby hit usage limits quickly, while Tyler Willis, Alex Vulkoff, and Fabio Jonathan argue Fable is not crazily token-hungry and can be cheaper due to one-shot solutions.
  • Ali K Miller says Fable 5 transformed her weekends, calling it an 'actual leap' that autonomously solved a tricky MBA-level word math problem with zero babysitting.
  • Riley Brown one-shot a Swift app replicating Replit mobile with four prompts, prompting debates about AGI claims versus the infrastructure work behind real companies.
  • Stripe reported Fable 5 compressed months of engineering into days, performing a codebase-wide migration on a 50-million-line Ruby project in a day versus a team’s two months.
  • Todd Saunders described Fable building a fully working product feature in real-time during a customer call, creating an 'autonomous looped building' workflow.
  • Felix Ryberg argues Fable 5 initiates a third AI era: moving from asking questions to assigning tasks, and now to giving responsibilities like autonomous loops monitoring crash reports.
  • Nate B. Jones argues the critical new skill is 'task imagination' - conceiving projects that leverage models capable of running for days, which most users currently lack.
Also from this episode: (4)

Models (3)

  • API pricing for Fable 5 is $10 per million input tokens and $50 per million output tokens, double Opus’s cost but lower than some expected. Mythos preview within Project Glasswing costs more than double.
  • Anthropic will remove Fable from subscription plans on June 23rd, moving to usage-based pricing. Whittemore calls this evidence of a 'firmly usage-based pricing paradigm'.
  • Whittemore notes Fable 5 can push back and disagree strategically, then update its position without fully collapsing, making AI-backed ideation more valuable.

Safety (1)

  • Fable 5 has strict guardrails, automatically routing queries about cybersecurity, biology, chemistry, or distillation to Opus 48. Anthropic says 95% of sessions don’t trigger a fallback but is 'hardcore' about biology/chemistry filters.
#Models#Coding#Agents#Enterprise#AI Infrastructure

AI Optimism vs. AI Pessimism • Jul 14

  • Nathaniel Whittemore states the Trump administration's AI executive order was signed after being pulled twice, with the final version making safety testing voluntary and reducing the pre-release review period from 90 to 30 days.
  • Whittemore says David Sacks intervened to stop the initial signing, and the NSA is assigned primary responsibility for testing under the order with support from cyber and defense agencies.
  • Dean Ball argues the executive order is a win for safety advocates and tees up a future licensing regime, while Steve Bannon predicts mandatory testing will be implemented within months.
  • Nathaniel Whittemore notes Anthropic expanded its Project Glasswing, adding 150 new partners across 15 countries in sectors like energy, healthcare, and communications to test its Mythos model.
  • Whittemore cites The Information's report that Mythos testers are spending millions on tokens, with Anthropic subsidizing costs, and firms are planning budgets to build strategies around the model.
  • Whittemore highlights OpenAI's report showing Codex growth, noting it has 5 million weekly active users and non-technical knowledge workers are adopting it three times faster than developers.
  • Nathaniel Whittemore says OpenAI identifies three frictions in knowledge work: finding inputs across sprawling systems, information coordination costs, and approvals and verifications.
  • Whittemore states 72% of Codex knowledge workers produce artifacts weekly, 41% do research, 27% perform data analysis, and 15% implement business workflows, with half now running parallel tasks.
  • Nathaniel Whittemore describes Codex's new features: Annotations for precise document interaction, role-specific plugins bundling 62 apps and 110 skills, and Sites for turning artifacts into shareable web apps.
Also from this episode: (4)

Chips (1)

  • Nathaniel Whittemore reports SK Hynix plans to double memory chip capacity by the decade's end to address AI-driven shortages, with Chairman Chey Tae-won stating the shortage could last until 2030.

Enterprise (3)

  • Nathaniel Whittemore notes Uber has instituted a $1,500 monthly token spending cap for employees, signaling cost management as a key vector for the next wave of enterprise AI.
  • Whittemore says Microsoft announced seven new AI models, including the 1-trillion parameter MAI Thinking 1, positioning them as part of a cost optimization strategy for enterprise adoption.
  • Nathaniel Whittemore cites Mustafa Suleyman stating that when tuned for McKinsey's tasks, MAI models outperformed GPT-5.5 on quality while being ten times lower on cost.
#Regulation#Big Tech#Models#Agents#Enterprise
Nerd Snipe with Theo and Ben
Nerd Snipe with Theo and Ben

5 different models dropped last week & the GPT-5.6 usage limits are brutal • Jul 14

  • Ben says OpenAI's subagent v2 implementation incorrectly copies the entire message history to spawned agents, wasting cache writes and increasing token counts.
  • Ben describes Grok Build's CLI as a polished, fast harness clearly inspired by Cursor and Claude Code, benefiting from XAI's engineering discipline.
  • Theo argues OpenAI's Codex branding was confused, and folding it into the ChatGPT super app with hidden URLs like chatgbt.com/codex hurts discoverability.
  • Ben says Apple's lawsuit against OpenAI alleges hardware chief Tan Tan downloaded confidential files after poaching 40 Apple employees.
  • Ben states the OpenAI employee in the Apple lawsuit bragged about retaining access to Apple's shared folders.
  • Theo argues GPT-5.6 Luna is best for programmatic calls like permission checks or title generation, not as a subagent.
  • Theo notes Codex desktop now opens the ChatGPT desktop app, and chat history is folded into a popup within the new interface.
Also from this episode: (5)

Models (4)

  • Ben argues OpenAI's Ultra mode is a token furnace because it forces subagents to use the inefficient Max reasoning level.
  • Ben claims GPT-5.6 Soul can consume 10-15% of a five-hour usage window in a single prompt without fast mode, up to 40% with fast mode enabled.
  • Ben and Theo agree Grok 4.5 is the first non-frontier lab model that handles complex, multi-step developer tasks without getting lost, competing with OpenAI and Anthropic.
  • Theo says Grok 4.5's price makes it a compelling option, noting his X Premium subscription includes $200 monthly credits.

Open Source (1)

  • Ben says OpenAI's Terra release has been overshadowed by Soul, making it a forgotten model.
#Models#Coding#Big Tech#Startups
Moonshots with Peter Diamandis
Moonshots with Peter Diamandis

The AI Duopoly Is Over: Grok 4. 5 , GPT-5 . 6 , and Muse Spark in One Week | #270 • Jul 13

  • OpenAI released GPT-5.6 on July 9th, a family of models including Sol, Tara, Luna, and Ultramode. The release moves OpenAI towards recursive self-improvement, using the high-end Sol model to post-train the lower-end Luna.
  • Elon Musk announced Grok 4.5 on July 8th. Meta released Muse Spark on July 9th, positioning it within its apps like WhatsApp and Messenger. Alex Gleas argues the frontier is no longer a duopoly; four American labs now operate at the optimal frontier.
  • Dave London says OpenAI's pivot from consumer to enterprise is evident; Chat GPT Work is a cargo-cult imitation of Anthropic's Claude app, focusing on revenue per token from code generation.
  • Peter Diamandis believes distribution is the new moat for AI models. Meta has 3.56 billion daily users, Google reaches 2.5 billion globally, and OpenAI has a billion monthly active users.
  • Alex Gleas sees an intelligence-polarized future: cheap, embedded AI on wearables versus high-cost frontier models in data centers driving scientific discovery. He argues profits from the high end fund the compute clusters.
  • AI-generated performer Tilly Norwood was cast as the lead in the feature film 'Misaligned'. The Screen Actors Guild condemned the casting, arguing it devalues human artistry.
Also from this episode: (6)

Big Tech (1)

  • Apple sued OpenAI for trade secret theft, alleging OpenAI stole confidential files and code names to build its AI hardware with Johnny Ive. Salim Mayel thinks Apple filed in Northern California because it is desperate to slow competitors while catching up.

Space (2)

  • Elon Musk tweeted SpaceX will be worth more than Earth if it accomplishes its goals. Peter Diamandis notes Earth's material wealth is about $600 trillion; adding financial assets brings it to $1.7 quadrillion.
  • China's Long March 10B booster successfully landed on July 10th, a first for China. SpaceX's Falcon 9 has achieved 580 booster reflights; booster 1067 flew its 36th mission on July 11th.

Robotics (1)

  • 1X Neo unveiled a redesigned hand with 25 degrees of freedom, tendon-driven and waterproof. Burn Borneck plans to produce 10,000 units in 2026.

Safety (1)

  • Illinois Governor J.B. Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, requiring frontier labs earning over $500 million annually to conduct third-party safety audits and report incidents within 72 hours.

Politics (1)

  • A new EU law mandates infrared driver monitoring cameras in all new cars and vans sold from July 7th. The system triggers alerts if a driver glances away for more than 3.5 seconds above 31 mph. Brussels claims it will save 25,000 lives by 2038.
#Models#Big Tech#Space#Robotics#Startups

Related Stories