Cursor, a third-party tool with no privileged access, extracts more performance from Anthropic's Opus 4.7 model than Anthropic's own Claude Code harness does.
Codex beats Claude Code, Stripe launched the AI bank & SpaceX targets a $7.5T valuation | GMP EP11
Cursor — a third-party tool with no privileged access — extracts more performance from Anthropic's own Opus 4.7 than Anthropic's own Claude Code does.
May 1, 2026
1:24:33
Difficulty: Intermediate
Played
God Mode Podcast
Codex beats Claude Code, Stripe launched the AI bank & SpaceX targets a $7.5T valuation | GMP EP11
Cursor — a third-party tool with no privileged access — extracts more performance from Anthropic's own Opus 4.7 than Anthropic's own Claude Code does.
May 1, 2026
1:24:33
Difficulty: Intermediate
Played
Chapter
No indexed bits in this chapter.
Snapshots ()
Stats
Episode stats
Insight Overview
insights
chapters
Insight distribution
Sub-Categories
Speaker breakdown
Talk Time
Key Quotes ()
This episode
Cast
-
Referenced as having bid $60 billion for Cursor and as having a new $7.5T SpaceX pay package requiring 100 TW of compute and a Mars colony.
-
Former US AI czar and All In podcast co-host, discussed as increasingly vocal in criticising Anthropic's alleged fear-mongering around the Mythos model.
-
AI-powered IDE and coding harness discussed extensively as outperforming Claude Code on Opus 4.7 and launching a new SDK.
-
AI lab discussed as maker of Claude and Opus 4.7, whose own harness (Claude Code) is being outperformed by third-party tools, and was excluded from the Department of War AI partner list.
-
AI lab discussed as maker of GPT 5.5 and Codex, building an AI-first phone, and named as a Department of War AI partner.
-
Discussed as approving a $7.5T valuation pay package for Elon Musk with Mars colony and compute clauses, and as a Department of War AI partner.
-
Discussed as launching Stripe Treasury — an AI-native banking product with MCP integration, 2% cashback, and multi-currency support.
-
Discussed as facing an existential threat from OpenAI's planned AI-first phone; Tim Cook stepped down and was replaced by John Turnis.
-
Chinese AI lab whose new model is reportedly 90-95% cheaper than GPT 5.5, making it a compelling option for routine agentic tasks.
-
Elon Musk's brain-computer interface company, mentioned as available for investment at a $43 billion valuation on pre-IPO crypto platforms.
-
OpenAI's coding product discussed as becoming an everything-app with browser and HeyGen video integration, seen as a direct competitor to Claude Code.
-
Anthropic's own CLI coding harness, discussed as being outperformed by Cursor on Anthropic's own Opus 4.7 model.
-
XAI's new model launched at roughly one-quarter the pricing of competitors and took #1 on Bridgebench for lowest hallucination rate.
-
Anthropic's unreleased cybersecurity AI model, discussed as possibly comparable to GPT 5.5's security capabilities and the basis for Claude Security.
-
Announced shift from flat-rate subscriptions to usage-based billing starting June 1, 2026, seen as a signal of the end of the AI subscription era.
-
Cloud-based IDE discussed as an alternative to Cursor, used by Luca and noted for active model-switching in its back-end.
-
Used as a blockchain analogy for AI model labs — the base layer that provides infrastructure but captures less value than the abstraction layer above it.
-
A web-scraping API platform cited as a key example of proprietary data infrastructure that AI agents will increasingly depend on.
-
Anthropic's new public-beta cybersecurity product for enterprise customers, suspected to be based on the Mythos model.
-
Solana's dominant wallet app used as an analogy for the AI abstraction layer (Cursor), which captures more value and brand equity than the underlying infrastructure.
This episode
Claims & Sources
4 / 20 cited (20%)
Factual claims made this episode, and whether a source was named.
GPT 5.5 performs better inside Cursor than inside OpenAI's own Codex environment.
Cursor was founded in 2022.
Elon Musk placed a bid valuing Cursor at $60 billion.
Grok 4.3 is priced at $1.25 per million input tokens and $2.50 per million output tokens, compared to Anthropic's $5/$25 and OpenAI's $5/$30.
Grok 4.3 claimed the number-one position on Bridgebench with the lowest hallucination rate of any available model.
GPT 5.5 became the most expensive AI model per token, overtaking Opus 4.7.
Despite higher per-token pricing, GPT 5.5 uses fewer tokens to solve the same problem as Opus 4.7, making it cheaper in effective cost.
GitHub Copilot announced a shift from flat-rate subscriptions to usage-based billing effective June 1, 2026.
The latest DeepSeek model is 90 to 95% cheaper than GPT 5.5 and comparable Claude models.
Global data-centre compute capacity currently sits at approximately 0.1 terawatts.
Elon Musk's SpaceX pay package requires hitting a $7.5 trillion valuation, deploying 100 terawatts of compute in space, and establishing a permanent Mars colony of at least 1 million people.
The Human Genome Project took 13 years to sequence the human genome and only achieved 92% completion.
The US Department of War named seven AI partners: SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, and AWS — excluding Anthropic.
Over 1.3 million Department of War personnel have used the AI platform, generating tens of millions of prompts and deploying hundreds of thousands of agents in only five months.
Google invested $40 billion into Anthropic.
OpenAI is reportedly building an AI-first phone in collaboration with Qualcomm, targeting a 2028 launch.
The 19-year-old creator of CalAI launched a stealth startup that reached $177,000 MRR within one month of launch.
Neuralink shares were available for investment at a $43 billion implied valuation on a crypto-based pre-IPO platform.
Tim Cook stepped down as Apple CEO, with John Turnis set to take over effective September 2026.
TL;DR
Three AI builders break down the biggest week in AI yet: Cursor's third-party harness outperforms Anthropic's own Claude Code on Opus 4.7, Codex launches as an everything-app with HeyGen video and browser baked in, Grok 4.3 drops at one-quarter competitor pricing, and GitHub Copilot kills flat-rate subscriptions. Stripe Treasury turns Stripe into an AI-native bank, SpaceX approves Elon's $7.5T pay package with a Mars colony clause, and the Department of War picks seven AI partners — pointedly excluding Anthropic. Key takeaway: the abstraction layer (Cursor, Phantom) captures more value than the base-layer infrastructure (model labs, blockchains).
#AI coding harnesses
#Cursor vs Claude Code
#model pricing wars
#abstraction layer value accrual
#SpaceX Mars colony
#Stripe Treasury MCP
#Department of War AI contracts
#Anthropic exclusion
#Grok 4.3 benchmark
#GitHub Copilot usage billing
#vibe coding
#whole genome sequencing
#agentic workflows
#AI phone OpenAI
#compute constraints
#Cursor
#Claude Code
#Codex
#Grok 4.3
#Stripe Treasury
#SpaceX
#Elon Musk
#Anthropic
#OpenAI
#harnesses
#abstraction layer
#GitHub Copilot
#usage-based billing
#Department of War
#AI agents
#Mythos
#genome sequencing
#MCP
#AI valuations
2 minute taster
Cursor Outperforms Anthropic on Anthropic's Own Model
Cursor extracts more performance from Opus 4.7 than Anthropic's own Claude Code does — despite having no privileged access. The harness is a different discipline from model training, and Cursor's years of multi-model developer experience give it an edge the labs can't easily replicate.
The Abstraction Layer Thesis: Phantom Prints, Solana Doesn't
In crypto, Phantom wallet made more money than Solana itself. The same dynamic is playing out in AI: Cursor and other abstraction layers will capture more value than the model labs running the underlying compute. Every tool call is a potential revenue event.
SpaceX's $7.5T Pay Package: 100 Terawatts and a Mars Colony
Elon Musk's SpaceX compensation package only vests if SpaceX hits a $7.5 trillion valuation, launches 100 terawatts of compute into space (1,000 times today's global data-centre capacity of 0.1 TW), and establishes a Mars colony of 1 million people. People laughed at the Tesla package too.
Look closer
Three builders discuss the biggest AI news of the week: Cursor outperforming Claude Code on Anthropic's own models, Codex becoming an everything-app with HeyGen integration, Grok 4.3 dropping at a fraction of competitors' prices, GitHub Copilot moving to usage-based billing, Stripe Treasury launching with MCP support, SpaceX's $7.5T pay package for Elon, and the Department of War naming seven AI partners while notably excluding Anthropic.
Chapter list
- Harness
- In AI development, the software layer (prompts, memory management, tooling) that wraps a base model to extract optimal performance for a specific use case — distinct from the model itself.
- MCP (Model Context Protocol)
- An open standard launched by Anthropic that lets AI models connect to external tools and data sources in a standardised way.
- Agentic workflow
- An AI-driven process where one or more models autonomously plan and execute multi-step tasks, often calling external tools, without continuous human input.
- IDE (Integrated Development Environment)
- Software that bundles a code editor, debugger, and build tools into one interface; in AI coding contexts, products like Cursor and Replit act as AI-enhanced IDEs.
- Context window
- The maximum amount of text (tokens) a language model can process in a single interaction; a larger window allows the model to 'remember' more of a conversation or codebase.
- Vibe coding
- Casual, natural-language-driven software development where a user describes what they want and an AI agent writes the code, lowering the barrier to entry.
- MRR (Monthly Recurring Revenue)
- A normalised measure of predictable monthly revenue from subscriptions or recurring contracts; a key SaaS health metric.
- TAM (Total Addressable Market)
- The total revenue opportunity available for a product or service if it achieved 100% market share.
- Inference
- The process of running a trained AI model to generate outputs; inference costs are the compute expenses incurred each time a model responds to a query.
- Supervoting shares
- A class of stock that carries more votes per share than ordinary shares, allowing a holder to retain outsized control of a company.
- Whole genome sequencing
- A lab process that reads the complete DNA sequence of an organism's genome, providing far more detail than standard genetic tests.
- Terawatt
- One trillion watts of power; used here as a measure of global data-centre compute capacity — the world currently sits at roughly 0.1 terawatts.
- Hallucination (AI)
- When an AI model confidently generates factually incorrect or fabricated information, presenting it as true.
- Agnostic (provider-agnostic)
- Describing a tool or workflow that is not tied to a single vendor's models or services, allowing easy switching between providers.
- Abundance
- As used by Elon Musk, a civilisational goal of having more than enough energy, compute, and resources for all humanity, driven by space expansion and AI.
- OAuth
- An open standard for access-delegation that lets users authorise third-party apps to access their accounts without sharing passwords; referenced here in the context of OpenClaw losing OAuth access.
- Abstraction layer
- A software tier that hides underlying complexity, letting users or developers interact with a simplified interface rather than raw infrastructure — analogous here to wallets in crypto sitting above blockchains.
- CICD pipeline
- Continuous Integration / Continuous Deployment: an automated software workflow that builds, tests, and ships code changes with minimal human intervention.
- Phantom (crypto wallet)
- A popular non-custodial wallet for the Solana blockchain, used here as an analogy for the high-value abstraction layer that captures fees from end-users.
- Comptue-constrained
- A state where demand for AI model usage outstrips available GPU/compute supply, forcing providers to impose rate limits or raise prices.
