OpenAI targets the banks stuck on COBOL

Hi there, this is your daily ☕️ AIpresso.

OpenAI's new model migrates old COBOL code LINK

• OpenAI says its latest models are now substantially better at reading and migrating legacy code in languages like COBOL, Fortran and Visual Basic, targeting the roughly 40% of banks still running COBOL systems facing a shrinking pool of programmers.
• On OpenAI's new MultiLangBench suite, which tasks models with building SQLite-compatible database engines across 20 languages, GPT-6 Astra scored an 87.4% average pass rate, more than triple GPT-5.6 Sol's, though COBOL stayed hardest.
• On COBOL-to-C++/Java/Kotlin/Python/Rust migrations that must preserve rounding and missing-record behavior, GPT-6.1 Sol hit 97.9% accuracy at high reasoning effort, dropping to 86.6% at low effort, though OpenAI warns production use still needs expert review against business rules.

New diffusion AI beats GPT-6's fast model LINK

• Inception released Mercury 3, a diffusion language model it says beats speed-optimized rivals GPT-6 Luna and Claude Haiku 5.5 on quality benchmarks while running four to eight times faster.
• The model generates over 1,000 tokens per second on standard NVIDIA GPUs and supports a 1M-token context, while being over 40x larger than Mercury 1 yet holding the same speed.
• It's available via the Inception API, Baseten, OpenRouter, and Vercel, and is OpenAI API-compatible so developers can swap it in without major code changes.

Gemini AI finds new fusion field solutions LINK

• Google's Gemini 4 Argon helped researchers extend exact mathematical solutions for the magnetic fields that confine fusion plasma, building on September 2026 work that overturned a conjecture physicist Harold Grad proposed back in 1967.
• Running through Google's Antigravity environment, multiple agents explored Matt Landreman's earlier results and, within hours, found ways to break a geometric symmetry restriction, yielding new configurations researchers could verify as reference cases for magnetic-equilibrium software.
• The work gives teams additional exact solutions to test plasma-modeling code against, not a path to commercial fusion, and the AI candidates still required manual derivation checks to confirm they were mathematically consistent.

Kling AI 4.0 adds stereo audio LINK

• Kling AI 4.0 is out, adding two-channel stereo audio and improved lip sync across more than ten languages including German, English, Japanese, Korean, and Spanish, alongside higher resolution and more precise controls.
• The model generates up to 30 seconds in a single pass, accepts up to ten keyframe images, and comprehends prompts up to 8,000 tokens, with its "Omni Reference" system taking up to 15 reference assets per generation.
• Kling 4.0 now works through MCP and CLI for agent-driven automated workflows, and a lower-cost Flash variant tops out at 720p and 20 seconds, though API pricing and support are still listed as "coming soon."

Cloudflare's Clairic reads docs cheaply LINK

• Cloudflare has released Clairic, a model for pulling structured data from documents that it says runs faster and cheaper than LLMs, now in preview on Workers AI with open weights to follow.
• Developers name the fields they want and Clairic returns matching values with a confidence score in a single pass rather than generating token by token, returning "null" for missing values and handling simple calculations and forecasts.
• In a sample complaint-handling chain with decision model Clef and GLM-5.3, the pipeline runs in 19.3 seconds versus 61.7 for GLM-5.3 alone at nearly double the cost, though Clairic can't yet reliably handle longer calculation chains.

Strategies & Tactics

> Cutting our inference bill by 68% by being obsessive about optimization: GitLab cut its AI code-reviewer cost 68% by sharing cached prompt text across repositories, trimming wasted model calls, and switching to cheaper, more efficient models.

> The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do: Keep the AI agent that routes tasks free of any reasoning, so failures stay contained and specialist agents handle each domain reliably.

> Accelerating Mamba-2 with High-Performance Forward Pass Kernels in NVIDIA CUDA Tile: Writing the Mamba-2 sequence model's forward pass in Nvidia's Python kernel language cuTile made it 1.14x to 1.82x faster than a comparable Triton version.

> One Search, Two Layers: Compiler Controls Inside PyTorch Helion using NVIDIA CompileIQ: Letting PyTorch's kernel tuner search compiler-level assembly settings alongside its usual choices squeezes extra speed from GPU code already considered fully optimized.

> MPS in PyTorch 2.15: PyTorch 2.15 speeds up local AI models on Apple Macs, adding memory-saving formats and faster decoding so Macs become a practical place to run and train models.

Other news you might like

• Nscale and Anthropic are reportedly in talks for yet another $45 billion compute deal LINK

• China’s tech giants drop $35 billion on export-compliant Nvidia RTX Pro 5500 workstation GPU orders to power AI ambitions, according to reports LINK

• DeepSeek made its name with cheap AI models. Now it’s undercutting search API rivals, too. LINK

• Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’ LINK

• fal launches FLUX 3 Max, claiming 22× faster AI video generation than Seedance 2.5 LINK

• Anthropic has nearly stopped hiring junior staff because AI agents now work six times as many hours LINK

• Replit is no longer working with Anthropic, CEO says LINK

Trending tools

Backdrop: build specialized AI coworkers connected to Slack, Asana, Hubspot and 1,500+ tools, with shared context and audit trails for research and reporting. LINK

face2social: upload a photo to find matching faces across Instagram, Facebook, LinkedIn, TikTok, and X, helping you verify identities and spot scammers. LINK

Train with Pasta: generates custom workouts from voice or text requests, swaps exercises on command, converts YouTube links into sessions, and keeps health data on-device. LINK

Dreamscale Labs: runs frontier robotics AI models like VLAs and WAMs through low-latency cloud inference, letting teams skip costly GPUs for on-demand access. LINK

SaidTrace: records audio on iPhone, Watch, or recorder, transcribes it on your Mac with your chosen AI, and links answers back to the source. LINK

Stage: publish prototypes from Figma, Claude, Cursor or GitHub, collect pinned feedback, auto-fix with Claude, and file Linear or Jira tickets on sign-off. LINK

Trending papers & reports

Robot vision planning lets robots figure out where to look and mentally picture what they'll see before grabbing objects, hitting ~94% success in real trials even when clutter or odd angles hide things. LINK

Safety guardrails for learning agents let reinforcement learning systems enforce complex rules during training without ever violating them, with mathematical proof the agent stays safe and still learns efficiently. LINK

Recommendation sets give high-stakes decision makers a spread of diverse, well-reasoned options each paired with a plain-language rationale instead of one ranked list, scoring highest on quality across deal-making, customer targeting, and content picks. LINK

Mislabeled training data leaves a detectable geometric fingerprint inside a neural network's internal representations, giving teams a way to spot corrupted labels that standard training accuracy completely misses. LINK

Explainable reinforcement learning gets a clear map, organizing over 250 studies by what each method explains and how, helping teams pick tools to understand why trained decision-making agents act as they do. LINK

See you tomorrow for a new dose of ☕️ AIpresso!