AI News — June 12, 2026 | Talking Fingers

Model releases

[2026-06-11] SpaceX / xAI — priced its IPO at $135 per share (555,555,555 shares), raising $75 billion — the largest IPO on record — at a valuation of about $1.75 trillion, with trading debuting June 12. Because SpaceX absorbed xAI in this year’s merger, this makes xAI the first frontier lab to reach public markets — landing days after OpenAI’s June 8 confidential S-1 and Anthropic’s June 1 filing, both of which analysts expect to list above $1 trillion. A frontier lab’s training spend, Grok revenue, and compute commitments now flow into quarterly public disclosures — the visibility the sealed Anthropic and OpenAI S-1s don’t yet provide. (source, source)

Being first onto public markets installs an earnings clock the still-private OpenAI and Anthropic won’t hear yet — a lab that spent its private years optimizing for raw capability now answers to shareholders pricing every compute commitment against margin, the pressure most likely to bend Grok’s roadmap and pricing from here.

Coding agents

[2026-06-10] Cursor — shipped a major Bugbot upgrade: reviews are over 3× faster (average ~90 seconds, down from ~5 minutes), 22% cheaper per run, and find 10% more bugs per review (0.62, up from 0.56). Bugbot and Security Review can now run before pushing via /review (or /review-bugbot and /review-security directly), and Bugbot can be configured to review only what changed since the last review. Together with Copilot CLI’s /security-review from the same day, pre-push agent review in the terminal is converging into a standard surface across vendors. (official)

Reviewing only what changed since the last pass is what makes pre-push review survivable on an active branch — a full re-scan on every iteration is exactly the friction that trains developers to start clicking past the bot.

For Software Developers: On a branch you’ve already pushed a few times, running /review before the next push now scans just the latest delta in around 90 seconds, so it slots into your pre-push routine the way a linter does — you catch the bug while the diff is small and still in your head, instead of reading Bugbot’s comment after the push once you’ve moved on to the next thing.
[2026-06-11] GitHub Copilot — Copilot CLI added a /settings slash command: a single schema-driven dialog that consolidates scattered commands (/theme, /streamer-mode, /experimental) and options that previously required hand-editing the settings file; it supports a full-screen browser, inline sets (/settings autoUpdate true), and per-setting resets. Small on its own, but it signals CLI agents maturing from flag sprawl toward discoverable configuration. (official)

For a tool whose options had been accreting across one-off commands and a hand-edited settings file, one searchable dialog is mostly an onboarding and discoverability win — the settings you didn’t know existed become the ones you can find without grepping the docs or copying someone’s dotfiles.

AI cost tracking & telemetry

[2026-06-11] AWS — launched AWS FinOps Agent in public preview at FinOps X’s closing day: an agentic system built on Amazon Bedrock that answers cost questions in natural language, investigates cost anomalies to root cause, surfaces optimization opportunities, and runs recurring FinOps workflows on a schedule, routing alerts into Slack and Jira. It runs in US East (N. Virginia), manages costs across regions and accounts from a management account, and is free during the preview. The anomaly-to-root-cause loop is the part FinOps keynotes spent the week saying teams lack for AI spend. (official, official, source)

Routing findings into Slack and Jira and running on a schedule is the tell that this is meant to live inside an existing FinOps workflow rather than be queried ad hoc — the anomaly investigation arrives as a ticket in the queue you already work, not a dashboard someone has to remember to open.

For Solution Architects: If you own a multi-account org under one management account, you can point the agent there and have a cost spike in any member account investigated to root cause and filed into Jira automatically — standing in for the manual Cost Explorer pass that usually starts only after someone notices the monthly bill jumped, and free to trial while it’s in preview.
[2026-06-12] Anthropic — the June 15 Agent SDK billing split is now officially documented, resolving the unconfirmed reports carried here since June 2: a Claude Help Center article confirms that from Monday, Claude Agent SDK, claude -p, and third-party Agent SDK apps move off plan usage limits onto a separate monthly per-user Agent SDK credit on Pro, Max, Team, and Enterprise plans. Credits must be claimed once (eligible users get an email with instructions before June 15), refresh each billing cycle, and can’t be pooled, transferred, or shared across an organization; interactive Claude Code, Cowork, and claude.ai are unaffected. Anyone running Claude in CI or via the SDK has three days to claim the credit — and decide whether to enable overflow usage credits — before automation hits the new meter. (official, source)

The clause most likely to catch teams off guard is the no-pooling rule: a per-user credit that can’t be shared or transferred means a shared CI identity or service account doesn’t get a pool the way a person does, so anyone running the SDK under a non-human login should confirm this week whether it even counts as an “eligible user” that can claim a credit.

Watch list

MiniMax M3 weights — deadline missed: the self-imposed “within 10 days” window closed June 11 with no weights or technical report on Hugging Face or GitHub. Meanwhile Artificial Analysis’ independent API-based evaluation scores M3 at 55 on its Intelligence Index v4 — just ahead of open-weights peers Kimi K2.6 (54) — under the pointed title “leading open weights model, once the weights are released”; the headline 59.0% SWE-Bench Pro remains self-reported. (source, source) (unconfirmed)

Worth watching now is the gap the Artificial Analysis headline points at — the 55 on the Intelligence Index is independently real, but for an open-weights pitch the weights are the product, so each day past the missed deadline turns a credible score into a model you still can’t actually run.
Gemini 3.5 Pro (late June): reportedly now in limited Vertex AI enterprise preview for select customers with broad GA expected late June — still no model card, public benchmarks, or confirmed pricing for the 2M-token context and Deep Think mode promised at I/O; Google has not announced the preview publicly. (source) (unconfirmed)

A limited Vertex preview with no model card is the awkward stage where the launch is real but still unverifiable — the signal to watch for is a public benchmark confirming the 2M-token context and Deep Think hold up against Fable 5’s bar, since an enterprise-only preview is nothing you can plan a build around yet.
xAI Grok V9-Medium (mid-June): the ~1.5-trillion-parameter coding-focused model is still expected this month — and with xAI now inside a public SpaceX as of today’s debut, its release lands under earnings-cycle scrutiny for the first time. (source) (unconfirmed)

What makes this one worth tracking past the parameter count is the timing the bullet flags — a coding-focused release would be xAI’s first real product proof point as a public company, so whether V9-Medium genuinely challenges the coding leaders or merely ships is now a signal the market reads, not just developers.
Fable 5 plan window closes June 22: ten days left of included Fable 5 on Pro/Max/Team/Enterprise; from June 23 it requires usage credits — the second new Anthropic meter this month after Monday’s Agent SDK split. (official)

Ten days out, this is the more concrete of the two Anthropic meters because both the date and the consequence are fixed — included today, usage-credit-metered from June 23 — so the open variable to watch is the credit pricing Anthropic still hasn’t published, which decides whether keeping Fable 5 in your daily loop past the 22nd stays affordable.