AI Forecast Tracker
Week 25 · June 15 – June 21, 2026
high signal

The Frontier, Recalled

The most capable model anyone could rent lasted three days. Anthropic shipped Claude Fable 5 on a Tuesday; that Friday the US government ordered it pulled worldwide — and the trigger was the exact skill the launch had celebrated, the model reading a codebase and fixing its flaws. It was the first time Washington has reached into a live commercial frontier model and switched it off, and at week's end it was still dark. Around that hole the rest of the industry carried on as if nothing were wrong: Salesforce paid $3.6 billion for an AI agent and Microsoft opened Office's memory to everyone else's, the clearest sign yet that the agent layer is consolidating by acquisition. PwC priced an AI skill at a 62 percent wage premium while the entry-level rung kept shrinking. And nine days before the bet that the AI boom bursts by 2026 comes due, DeepSeek raised $7.4 billion, SpaceX bought Cursor for $60 billion, and data-center spending stayed north of a trillion — the boom is not bursting. Brussels, meanwhile, postponed its own crackdown to 2027 the same week Washington told its agencies to move faster.

7 predictions updated3 milestones25 companies refreshed

Key Developments

1

The most capable model ever made public was pulled by the US government three days after launch — over the coding skill everyone had just celebrated

On June 12 at 5:21pm ET, a US Commerce export-control directive ordered Anthropic to disable Claude Fable 5 and Mythos 5 for every customer worldwide — three days after Fable 5 had shipped as the first publicly accessible Mythos-class model (95 percent SWE-bench Verified, #1 on the Artificial Analysis Intelligence Index). The stated trigger was a narrow jailbreak: getting the model to read a specific codebase and fix its software flaws — the exact agentic-coding capability the launch had been celebrated for. Anthropic complied while disputing the action, dispatched staff to Washington, and as of June 20 the models were still offline, with prediction markets pricing roughly 57 percent odds of restoration before July 1. It is the first time a government has retroactively switched off a live commercial frontier model.

Confirms P-034 AI capabilities are advancing at a rate that outst...Challenges P-001 AI models will handle most aspects of software eng...Challenges
Counterpoint

The suspension is an access event, not a capability regression — Fable 5's 95 percent SWE-bench Verified is intact, just unreachable. Anthropic argues the jailbreak is "widely available from other models" and that applying the standard industry-wide "would essentially halt all new model deployments," and the market reads the outage as temporary (~57 percent restored before July 1). The durable signal is not lost capability but a new variable: frontier-model availability is now subject to a government switch, a structural risk that did not exist a week ago.

Source →
2

Salesforce paid $3.6 billion for an AI agent and Microsoft opened Office to everyone else's — the enterprise agent layer is consolidating by acquisition

Salesforce signed a $3.6 billion definitive agreement on June 15 to acquire Fin (the AI customer-service agent formerly known as Intercom), the largest acquisition aimed purely at agentic customer experience, folding it into Agentforce as the same release took multi-agent orchestration to general availability. A day later, Microsoft's Work IQ APIs went GA, opening the M365 semantic layer — email, files, meetings, people — to any third-party agent on consumption billing. ServiceNow's IT AI Specialists reached GA the same window and Workday made Gemini Enterprise the default model inside its Sana self-service agent. The common thread is that the agent layer is no longer just being built; it is being bought, platformed, and governed.

Confirms P-021 40% of enterprise applications will feature task-s...Confirms
Counterpoint

Buying the capability is not the same as making it stick — the Klarna Pattern. Deloitte's 2026 State of AI report found 42 percent of companies abandoned at least one AI initiative in 2025 (up from 17 percent in 2024) at an average $7.2 million sunk cost, and only about 12 percent of agent initiatives reach production at scale; a parallel survey put 72 percent of enterprises running agents but 60 percent without formal AI governance in place. A $3.6 billion purchase order and a GA flag are inputs, not outcomes — the materiality is whether these agents are still measured and running at year-end.

Source →
3

An AI skill now pays a 62 percent wage premium — and the bottom rung of the ladder keeps shrinking

PwC's 2026 Global AI Jobs Barometer, released June 15 from more than a billion job ads across 27 countries, found that roles demanding AI skills now pay a 62 percent wage premium (up from 57 percent a year earlier) and are growing 69 percent a year, while traditional entry-level postings have fallen 10 percent since 2019 and "seniorised" entry roles — junior titles that now demand senior judgment — rose 35 percent. The Economic Policy Institute's Class of 2026 analysis adds that 21 percent of firms have frozen entry-level hiring specifically because of AI and 47 percent expect to eliminate entry roles by 2027. Layoff trackers put 2026 tech cuts near 185,900 through June 19 — about 1,115 a day, with 56 percent of events citing AI. This is The Hollowing as a market price: the skill is rewarded, the apprenticeship is disappearing.

Confirms P-002 AI will disrupt 50% of entry-level white-collar jo...Confirms P-020 AI-exposed industries will see job losses of ~20,0...Confirms
Counterpoint

Attribution is still the soft spot. Deutsche Bank's "AI redundancy washing" thesis holds that a large share of cuts blamed on AI are pandemic-era overstaffing corrections dressed up, and an NBER study of thousands of workplaces found close to zero effect on earnings or hours at firms that actually adopted AI versus those that merely claimed to. Announced layoffs also realize at only 60-70 percent. The wage-premium data is the cleaner signal; the headcount-causation data is not yet settled.

Source →

What the Evidence Moved

P-02140% of enterprise applications will feature task-specific AI...

A single-week agentic cluster — Salesforce's $3.6B Fin acquisition plus multi-agent orchestration GA (June 15), Microsoft Work IQ APIs GA (June 16), ServiceNow IT AI Specialists GA, and Workday making Gemini Enterprise its Sana default. The platform layer is now consolidating by acquisition, not just shipping. At the extreme boundary (0.92); the open question stays P-022 (does it stick), not presence.

91%92% +1pp
P-022Over 40% of agentic AI projects will fail by 2027 due to uni...

Deloitte's 2026 State of AI report found 42 percent of companies abandoned at least one AI initiative in 2025 (up from 17 percent in 2024) at an average $7.2M sunk cost, with only ~12 percent of agent initiatives reaching production at scale; a parallel survey put 72 percent running agents but 60 percent without formal AI governance. The failure mechanism is now quantified at the abandonment level, not just the pilot level.

81%82% +1pp
P-002AI will disrupt 50% of entry-level white-collar jobs over 1-...

PwC's Barometer (June 15) priced an AI skill at a 62 percent wage premium while traditional entry-level postings fell 10 percent since 2019 and seniorised entry roles rose 35 percent; EPI reports 47 percent of firms expect to eliminate entry roles by 2027. Counter: NBER finds near-zero earnings/hours effect at firms that actually adopt AI, and the redundancy-washing caveat persists — small move up.

72%73% +1pp
P-020AI-exposed industries will see job losses of ~20,000 per mon...

Layoff trackers put 2026 tech cuts near 185,900 through June 19 (~1,115/day, nearly double 2025's pace) with about 56 percent of events citing AI, and TechCrunch framed ~40,000 May cuts as the highest single month in two years. Counter: stated-AI attribution overstates true automation and announced cuts realize at only 60-70 percent.

73%74% +1pp
P-034AI capabilities are advancing at a rate that outstrips the e...

The clearest evaluation-gap event yet — the US government suspended Fable 5 + Mythos 5 globally (June 12) after being shown a narrow jailbreak the model's safety evals had not flagged as deployment-blocking. A model judged safe to ship behaved unsafely enough in deployment to trigger a government takedown — the exact test-vs-deployment divergence the claim describes. NSPM-11 (June 5) simultaneously pushed faster defense AI adoption.

83%84% +1pp
P-017Annual data center CapEx will grow from $500B (2025) to $1.4...

Dell'Oro's Q1 2026 report confirmed global data-center capex above $1T annualized, with the top US providers committing $660-725B for 2026 (+77 percent YoY) and explicit 2H acceleration on the NVIDIA Rubin ramp — supply now runs ahead of ARK's $1.4T-by-2030 trajectory. The GPU-depreciation-vs-debt mismatch remains the channel that could break it before 2030.

59%60% +1pp
P-03130% of enterprises will abandon facial verification by 2026 ...

The European Parliament formally approved the Digital Omnibus amendments 423-57 (June 16), locking in the deferral of Annex III high-risk obligations to December 2027. The near-term regulatory forcing function that would have pushed enterprises off facial/biometric verification by year-end is now near-certain to be gone — further weakening the 30-percent-by-2026 case.

52%51% -1pp
P-012AI investment boom is a bubble likely to burst by 2026

Final weekly check, nine days to target_date 2026-06-30. The week ran opposite to the claim — Nasdaq +2.4 percent, DeepSeek's $7.4B raise, SpaceX's $60B Cursor purchase, >$1T data-center capex — and none of the five operational burst criteria are met. Resolves refuted at month-end absent a discrete shock.

5%5% 0pp

Company Impact

Anthropic

Score change

US Commerce export-control directive (June 12) forced a full global suspension of Fable 5 + Mythos 5 three days after launch, citing a narrow jailbreak; still offline as of June 20 with staff dispatched to Washington (~57 percent market odds of restoration before July 1). First demonstrated government veto of a deployed commercial frontier model — a disruption-risk category not priced at dr 2. dr 2->3. Score 9.1->8.85 (still very_positive). S-1 (June 1, $965B/$47B run-rate) and capability leadership intact.

Access suspension of Fable 5 and Mythos 5 — Anthropic · Anthropic suspends Fable 5 / Mythos 5 after export-control directive — Greenberg Traurig

8.8
See assessment →

Bristol-Myers Squibb

Score change

Strategic agreement with Anthropic (May 20) deploys Claude Enterprise as a shared agentic platform across 30,000+ employees in R&D, clinical development, manufacturing and commercial — a C-suite-sanctioned, enterprise-wide commitment moving adoption from Fast Follower toward Active Adopter. aam 6->7. Score 5.7->5.8 (stays neutral). The LOE cliff and absence of a Phase III AI-discovered candidate cap the move.

Bristol-Myers Squibb deepens AI with Anthropic — BioPharma Dive

5.8
See assessment →

Salesforce

Data refresh

Signed a $3.6B definitive agreement to acquire Fin (June 15, largest agentic CX acquisition to date; closes early CY2027) and took multi-agent orchestration GA the same day with Gemini 3.5 Flash embedded. Both deepen the agentic moat but are forward-looking — no adoption data yet, Fin does not consolidate until close — so the score holds 7.0 pending Q2 FY27 (devil's-advocate hold).

Salesforce acquires Fin for $3.6B — TechCrunch

7.0
See assessment →

Cursor (Anysphere)

Data refresh

SpaceX agreed to acquire the company for $60B in all-stock on June 16 (closing Q3 2026) to strengthen its xAI-based AI division — the largest acquisition of an AI-coding company to date. Whether the model-agnostic stance survives xAI ownership is the open question; score holds 8.1 pending close and integration signals (devil's-advocate hold — a forward-looking, partly-speculative event).

SpaceX to acquire Cursor for $60B — TechCrunch

8.1
See assessment →

Microsoft

Data refresh

Work IQ APIs went GA June 16, opening the M365 semantic layer (email, files, meetings, people) to any third-party agent on consumption billing — extending the platform moat beyond first-party Copilot. GitHub Copilot moved to usage-based billing June 1 (developer backlash a near-term friction). Dimensions saturated (are 9, md 9, aam 9); score holds 8.25.

Announcing the new Work IQ APIs — Microsoft 365 Blog

8.3
See assessment →

Alphabet / Google

Data refresh

Gemini Enterprise became the default model inside Workday's Sana self-service agent (HR/finance), while flagship Gemini 3.5 Pro stayed in limited Vertex preview as of June 21 — a second consecutive frontier slip, though the Flash substitute already beats the prior Pro generation. Distribution moat intact; score holds 8.0. A third slip would pressure disruption-risk.

Workday and Google integrate HR and finance AI agents into Gemini Enterprise — Futurum

8.0
See assessment →

Sources

Next issue drops Monday

Subscribe to get the briefing before the market opens.