OpenAI moves banking's reference data onto its own infrastructure
OpenAI launched ChatGPT for Financial Services on 10 September, a ChatGPT Work configuration scoped to investment banking and equity research, with Morgan Stanley and Evercore as design partners. The detail that matters is not the model. Premium datasets from Daloopa, PitchBook and LSEG News ship built in, and OpenAI states plainly that it indexes and hosts that data on its own infrastructure rather than reaching it through MCP connectors. The stated reason is retrieval quality and latency: hosting the corpus is what makes granular citations possible, so an analyst can trace an adjusted EBITDA figure back to the note it came from. Shared sign-in and entitlement integrations are being built with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva and Moody's.
The governance list is the standard enterprise one — SAML SSO, SCIM, role-based access, no training on business data by default, configurable retention, log export, separate workspaces to enforce information barriers. What the page does not contain is a single word about where any of this runs. No EU data residency, no regional availability, no DORA posture. For a European bank, that is the whole question: the controls describe who inside the firm may see the data, not which jurisdiction can compel its disclosure. Design partners are also not deployments, and both quoted firms describe collaboration rather than rollout. Availability is “eligible financial institutions” via a sales conversation, with no pricing and no GA date.
The cost story in frontier models is now verbosity, not price per token
Artificial Analysis published its independent evaluation of GPT-6 Astra on 9 September, and the finding cuts against the price list. Astra is priced at 2.5x GPT-5.6 Sol, $10/$50 per million tokens against Sol's $4/$20. It still ties Claude Fable 5.1 at the top of AA's Intelligence Index — both score 53, Astra at max effort against Fable 5.1 at max with fallback — at roughly 40% of the cost per task, $3.26 against $7.63. The mechanism is token efficiency: at maximum reasoning effort Astra spends 27,000 output tokens per task against Fable 5.1's 78,000 for the same score. Hallucination rate on AA-Omniscience fell from 92% to 51% at max effort, with accuracy up 4 points rather than traded away.
Two caveats belong with those numbers. The same shortness costs it elsewhere — Astra drops about 45 Elo on GDPval-AA v2 against its predecessor, and AA observes it taking 24 turns per task where GPT-5.6 Sol takes 45 and Claude Fable 5.1 and Opus 5 take 60. And at least three of the ranked entries on the Coding Agent Index are run in their own vendor's harness — Astra in Codex, Fable 5.1 in Claude Code, Muse Spark 1.3 in Muse Code — so those scores measure model plus scaffold, not model alone. The practical consequence for anyone buying inference is that a published rate card has stopped predicting a bill. Cost per completed task is a property of the model's behaviour, and you only see it after measuring.
A 30B Apache-2.0 model trained on preferences a machine graded
Korean lab Trillion Labs, with the Lunit Consortium, released Gravity-30B-A5B on 9 September — a sparse mixture-of-experts model, 29.56 billion total parameters with 5.34 billion active, 128K native context, Apache 2.0, ungated, LICENSE file present. It was not trained from scratch: the base model card records that the team took their earlier 16B model and used LLaMA-Pro-style block expansion to grow it from 28 layers to 52, then continued pretraining. The architecture is deliberately DeepSeek-V3-compatible, which matters for whether your serving stack can pick it up without custom work.
The interesting part is the alignment method. The final preference-tuning rounds used execution-labelled pairs: candidate tool calls were actually run against live tool environments and compared against the gold end state, with no model acting as judge. That removes one of the least defensible links in the usual pipeline. Two practical notes before anyone tries it — the instruct model is labelled a preview and requires trust_remote_code, which is a supply-chain decision in a regulated environment, and a same-day therapeutics variant, TxGravity-30B-A6B-Open, carries apache-2.0 in its card metadata but ships no LICENSE file at all.
Quick Hits
- Brussels proposes a European Innovation Act. — The Commission adopted the proposal on 9 September, framed around technological sovereignty. Two pillars: an EU framework for valuing intellectual property plus a digital IP marketplace, and a common procedure for R&D procurement with cross-border joint purchasing. A parallel proposal for a Council Recommendation would create a common reference framework for regulatory sandboxes, scoped in the Commission's own words to “areas not yet covered by EU legislation” — which on that scope should leave the AI Act's Article 57 sandboxes untouched. Both are proposals; neither is adopted law, and the euro figures attached to them are the Commission's own impact estimates.
- Transformers 5.17.0 breaks vision RoPE and adds a French encoder family. — The release landed 9 September with seven new model additions, including NeoMME from H Company — 260M and 800M multimodal encoders that, in the release's words, process “multilingual text tokens and raw image patches in a single bidirectional Transformer encoder, without a separately pretrained vision tower or causal language model”. The breaking change: 2D and 3D vision rotary embeddings were consolidated into one module, so custom vision models with their own grid interleaving need migrating. Note also that the new HYV4 support does not execute the MTP layers, so released checkpoints keep those weights for other runtimes.
- ONNX Runtime 1.30 changes three defaults at once. — The release of 10 September enables FP4 QMoE kernels by default in CUDA builds, cuts CUDA fpA-intB builds to a compact kernel set unless you opt back in, and makes CPU FP16 nodes without a matching kernel fall back to FP32. Go bindings for the C API are new. If you run CPU or edge inference, these are behaviour changes rather than additions — worth a regression pass before upgrading.
- The DOJ wants to know how Nvidia structured the Groq deal. — The New York Times reported on 9 September that antitrust officials sent Nvidia a formal request for information, examining whether the December transaction was arranged to avoid merger review. Nvidia billed it as a non-exclusive licensing agreement; Groq's founder and several key staff moved across. The price is reported rather than disclosed and outlets do not agree on it — Reuters carries the announced figure at $17bn, Bloomberg calls it a $20bn licence deal. A request for information is not a filed case.
- — Bloomberg , up from roughly 12 GW today, with AI-dedicated capacity growing to about a third of the total. The number excludes capacity rented from neoclouds, so it is not the same denominator as most published “Microsoft compute” figures. This is a report of plans, not a company announcement.