TechLook
France just put a serious competitor in the code generation ring. Mistral shipped Codestral, a 32K-context model trained on 80+ programming languages, and the benchmarks don't hedge. On RepoBench — a dataset that tests how well models complete code inside real repositories — Codestral outperforms GPT-4o. It also supports FIM (fill-in-the-middle) completion, which is how modern IDEs integrate code generation. For builders evaluating their stack, this isn't a "too early to tell" release.
Here's everything you need to know:
- Codestral scores above GPT-4o on RepoBench, a repository-level code completion benchmark — a more realistic test than isolated coding exercises
- Trained on 80+ programming languages, including niche ones outside the usual Python/JavaScript set
- 32,000 token context window covers most files and multi-file editing sessions
- Supports FIM (fill-in-the-middle) completion, the integration format used by VS Code and JetBrains IDEs
- Mistral positions it as a "global benchmark" model — meaning it's designed for general use, not just Mistral's own ecosystem
- Le Chat (Mistral's assistant) and Le Maire (their enterprise platform) remain separate products
- The API is available today via Mistral's platform and is being evaluated against Codex, Claude, and Gemini in active developer stacks
For indie hackers and SaaS builders who aren't married to a single provider, Codestral adds a genuine third option in the code generation layer. The RepoBench result against GPT-4o matters more than it might seem — real repository performance is a different test than synthetic benchmarks, and that's where production decisions actually get made. If you're building internal tooling or evaluating AI coding assistants, this week is a good time to re-run your own evals.
The Inference War Has a New Battlefield — and NVIDIA Just Planted Its Flag on It
The AI compute race used to be about training. That era is closing. NVIDIA's Blackwell refresh — the GB300 and GB300N — is explicitly built for the inference side of the workload, and the numbers tell you where the money is moving.
Here's everything you need to know:
- GB300N is the inference-optimised variant, drawing 1.4x the power of GB200 but delivering 2x the inference throughput
- Both GB300 and GB300N use the NVL72 rack configuration — 72 GPUs per rack — up from NVL36
- The GB300 (training variant) targets distributed training at scale, while the GB300N targets inference at scale
- NVIDIA is presenting this as an infrastructure upgrade path for hyperscalers already running Blackwell clusters
- The power efficiency delta is the real headline: 2x inference throughput for 1.4x power draw is a meaningful step change
- This follows AWS's Trainium3 announcement earlier this week, which also targets inference workloads with 2.4x performance over Trainium2
The inference cost curve is now the variable that defines AI business models. When inference gets cheaper and faster per unit of output, the economics of AI-native products improve across the board — lower cost-per-query for SaaS products, more viable freemium tiers, cheaper agentic workflows. If you're building AI features into a product, the trajectory on inference infrastructure matters as much as model capability. Watch this space.
GitHub Copilot Just Switched to Monthly Billing — and That's a Signal
GitHub quietly moved Copilot from a usage-based model to monthly subscription pricing. The shift matters not because of the price change itself, but because of what it says about how GitHub is thinking about the product's trajectory.
Here's everything you need to know:
- GitHub Copilot has transitioned from a pay-per-usage model to a fixed monthly subscription
- The change reduces billing unpredictability for engineering teams — a real friction point for budget owners
- It also signals GitHub is confident enough in Copilot's usage patterns to commit to predictable revenue
- For teams currently evaluating Copilot vs alternatives, fixed pricing changes the unit economics calculus
A usage-to-subscription shift is typically what happens when a product crosses from "new and experimental" to "reliable infrastructure." GitHub is essentially telling the market: we believe this is core enough to your workflow that you should budget for it, not treat it as variable spend. If you're a founder running engineering budget, this is a good moment to re-evaluate whether Copilot is factored into your headcount-equivalent math.
Perplexity Opened Deep Research to the API — Developers Finally Get Direct Access
Perplexity's Deep Research capability — the feature that lets you run multi-step web research tasks — is now available directly in the Perplexity API, with higher rate limits than the consumer product.
Here's everything you need to know:
- Deep Research is now accessible via API endpoint for all developers, not just enterprise customers
- API rate limits are higher than the consumer plan, making it viable for production workloads
- The feature handles multi-step research tasks — it browses, synthesises, and returns structured outputs
- Perplexity's web-grounded approach differentiates it from models that rely purely on training data
The API availability changes the build-vs-bridge decision for teams that were wrapping Perplexity in intermediary layers. If you've been running Perplexity queries through third-party tools or manual workarounds, this removes a friction point. For products that need current web-grounded reasoning — not just model knowledge — this is the piece that's been missing from the API-first stack.
Salesforce Put 200 Actions in AgentForce — and Added a Trigger-Based Agent Loop
AgentForce 2dx is Salesforce's most substantive update since launch. The headline: 200+ pre-built actions and a new event-driven architecture that lets agents respond to CRM changes without human triggers.
Here's everything you need to know:
- AgentForce 2dx ships with 200+ pre-built actions covering standard CRM workflows
- Event-driven agents can now trigger on CRM data changes (a record updated, a deal stage changed) — not just scheduled or manual runs
- 10,000 agent messages per month are included per organisation at no additional charge
- The enterprise angle is clear: this is for companies with established Salesforce deployments that want agentic automation without custom integration work
- The trigger-based model moves agents from "do a task when asked" to "monitor and act continuously" — a meaningful capability difference
For founders building on top of Salesforce or evaluating CRM integrations, the event-driven agent model is a preview of where enterprise automation is going. Not "run a workflow when a user clicks a button" but "run a workflow when data changes in a way that meets a condition." That shift — from reactive to proactive — is the core value proposition of AI agents in enterprise tooling, and Salesforce just made it configurable without code.
Anthropic Extended Claude 3.7 Sonnet's Thinking Budget — More Reasoning on Demand
Anthropic has made extended thinking capabilities a configurable feature in Claude 3.7 Sonnet, allowing developers to set higher thinking budgets for complex tasks.
Here's everything you need to know:
- Extended thinking allows Claude 3.7 Sonnet to allocate more compute to reasoning on complex problems
- The thinking budget is now configurable via the API — developers can tune it for task complexity
- The capability is designed for tasks where standard responses fall short — multi-step reasoning, complex analysis, technical problem-solving
- This follows the pattern Anthropic established with o3-mini and reasoning budgets in OpenAI's API
If you're building products that rely on Claude's reasoning for complex workflows — code review, document analysis, multi-step planning — the configurable thinking budget gives you a dial: spend more compute for harder problems, less for simpler ones. That's the granular control that production systems need.
⚡ Quick Hits
-
Microsoft AutoDev: Open-source agentic dev tool released under MIT license. Plans, searches, writes, and executes tests autonomously. Direct signal for tooling builders evaluating agentic coding infrastructure.
-
AWS Trainium3: Now live in us-east-2 with 2.4x performance improvement over Trainium2. The inference accelerator market is heating up — NVIDIA, AMD, and now AWS silicon all competing for the same workloads.
-
OpenAI Operator in EU: Operator — OpenAI's agent that acts in the browser on your behalf — is expanding to Europe. The question for builders: what does your product look like when your users' AI can do the browsing and form-filling for them?
-
Perplexity Sonar 32B: New large model variant in the Sonar family. Smaller than the original Sonar (which was 70B+), likely positioned for speed and cost optimisation at the expense of some capability.
-
AMD + Dell Enterprise AI: AMD and Dell announced a partnership targeting enterprise AI deployments. AMD's MI350 GPU line is the hardware backbone; Dell's enterprise relationships are the distribution. Another signal that the AI infrastructure market is being carved up by partnerships, not just individual vendors.
-
Cerebras IPO: Cerebras filed for IPO. The company behind the Wafer Scale Engine — a single wafer-sized chip for AI training — is going public. Worth watching for what the numbers tell us about the economics of dedicated AI silicon outside the hyperscaler ecosystem.
-
Samsung Knox AI: Samsung is adding Knox AI features for enterprise mobile device management. On-device AI capabilities for security and device management on Samsung hardware — a different angle on the enterprise AI story than the cloud-centric plays dominating headlines.
-
HP AI PC: HP announced new AI PC hardware targeting the commercial market. The AI PC category is consolidating around Copilot+ PC standards — another signal that on-device AI is moving from concept to procurement decision for enterprise IT buyers.
-
Canva + Google ChromeOS: Canva integration is expanding to ChromeOS with AI features powered by Gemini Nano. The design tool wars are extending into the OS layer — and Google is threading Canva's ecosystem into ChromeOS rather than building a competing product.
-
Meta AI Studio: Meta is opening AI character creation tools to developers via AI Studio. Build, deploy, and monetise AI personas. For founders building on the social/creator side of AI, this is the platform layer being built in real time.
-
GitHub Copilot Monthly: Already covered above — worth restating because the billing model shift is the story, not the price.
-
OpenAI o3-mini High behind $200/mo subscription: OpenAI's strongest reasoning model for the money is now effectively gated behind their highest subscription tier. The pricing stratification signal from the GPT-4.5 Instant default story is continuing — the free tier is becoming a lead-generation funnel for paid tiers, not a real product option.
Techlook — AI & tech signal for founders and builders.