The frontier moved in 30 days
Meta, xAI, OpenAI, and Anthropic each shipped a new flagship model between June 9 and July 9, 2026 — and two got pulled back by governments before the dust settled. Specs, benchmarks, pricing, funding, and the timeline, side by side.
Specs, benchmarks & pricing
Switch views — every number is footnoted where it's third-party sourced, conflicting, or simply undisclosed.
| Spec | META | xAI | OpenAI | Anthropic |
|---|---|---|---|---|
| Launched | Jul 9, 2026 | Jul 8, 2026 | Jul 9, 2026 (public) | Jun 9, 2026 (restored Jul 1) |
| Context window | 1M tokens | 500K tokens | ~1.5M — unconfirmed | 1M tokens / 128K out |
| Modality | Text, image, video, PDF, audio | Text + image in, text out | Not officially detailed | Text + high-res vision |
| Built for | Agentic + computer use | Agentic coding (w/ Cursor) | Agentic coding, cyber/bio-heavy | Long-horizon autonomous work |
| Reasoning | Multimodal reasoning, multi-agent orchestration | Always-on ‘high’ effort, non-disableable | Frontier agentic/reasoning tier | Always-on thinking, effort param low→max |
| Pricing (in / out per M) | $1.25 / $4.25 | $2.00 / $6.00 | $5.00 / $30.00 (Sol) | $10.00 / $50.00 |
| Access | Meta Model API (preview, US) | xAI API, Grok app, Cursor | API/preview only, gated | Claude API, Bedrock, Vertex |
State of the art
Widening the lens: Gemini 3.1 Pro and DeepSeek V4-Pro top the published benchmark scores outright — the four hero labs aren't the whole story. Every cell below is tagged by provenance: vendor-reported, independently confirmed, or not disclosed.
| Benchmark | Anthropic | OpenAI | xAI | Meta | DeepSeek | Qwen | |
|---|---|---|---|---|---|---|---|
| GPQA Diamond Anthropic's own launch post gives no number elsewhere — sources conflict on the exact figure. | 88.5 vendor | 94.3 vendor | not disclosed — | not disclosed — | not disclosed — | 90.1 vendor | 92.4 vendor |
| MMLU / MMLU-Pro | 91.2 vendor | not disclosed — | not disclosed — | not disclosed — | not disclosed — | 87.5 vendor | not disclosed — |
| SWE-bench Verified Anthropic's figure is the only one independently reproduced (vals.ai) — everyone else is vendor-reported or undisclosed. | 95.0 ✓ ind. | 80.6 vendor | not disclosed — | not disclosed — | not disclosed — | 80.6 vendor | 80.4 vendor |
| SWE-bench Pro Anthropic's 80.3% is contested — measured on its own scaffolding, not a neutral harness. | 80.3 vendor | not disclosed — | not disclosed — | 64.7 vendor | not disclosed — | 55 vendor | 60.6 vendor |
| ARC-AGI-2 Anthropic's ~12% may reference ARC-AGI-1, not -2 — treat with caution. | 12 vendor | 77.1 vendor | not disclosed — | not disclosed — | not disclosed — | 71.8 vendor | not disclosed — |
| Humanity's Last Exam Meta claims "state of the art" on HLE but publishes no number. | 53.3 ✓ ind. | 44.4 vendor | not disclosed — | not disclosed — | not disclosed — | 37.7 vendor | 41.4 vendor |
| LMArena Elo / rank | #1 at launch, suspended Jun 12 vendor | #3 overall vendor | not disclosed — | not disclosed — | not disclosed — | not disclosed — | not disclosed — |
| Artificial Analysis Index GPT-5.6 was too new for AA to score at publish time. | 60, #1 tracked vendor | not disclosed — | not disclosed — | 54, #4 of 168 vendor | 43, #22 of 551 vendor | not disclosed — | not disclosed — |
Six more models for context
- Context
- 1M tokens (2M claimed by some sources, unconfirmed)
- Pricing
- $2.00 / $12.00 per M
Benchmark leader — highest GPQA, ARC-AGI-2, MMLU of any tracked model
- Context
- 1M tokens
- Pricing
- $0.435 / $0.87 per M
Cheapest frontier-class model; open-weight; independently reviewed by NIST/CAISI
- Context
- 1M tokens
- Pricing
- $2.50 / $7.50 per M
Math/science challenger — claims to beat Western labs on HMMT, GPQA
- Context
- 262K tokens
- Pricing
- $0.60 / $2.50 per M
Best open-weight agentic/coding value; 1T MoE, 32B active
- Context
- 262K tokens
- Pricing
- $0.50 / $1.50 per M
Cheapest large open-weight model; enterprise-deployable, not benchmark-topping
- Context
- 1M tokens (10M for Scout variant)
- Pricing
- open-weight
Meta's last open-weight generation — superseded internally by proprietary Muse Spark; ~15 months dated vs. current frontier
Price vs. capability — the value frontier
Core four only; capability score is the editorial coding/agentic axis from the radar chart above.
Bottom-right is the best value: high capability at low price. Price = average of input/output per-M-token rates.
Model family tree
Every major release from GPT-1 to GPT-5.6, Claude 1 to Fable 5, Gemini 1.0 to 3.1, Grok 1 to 4.5, and Llama 1 through Meta's pivot to Muse Spark.
Restored globally after export controls lifted
Click any marker to see what shipped. 47 releases tracked across five labs, Feb 2018 – Jul 2026.
Funding, in full
OpenAI (founded 2015), Anthropic (2021), and xAI (2023) — from seed to nine-figure-and-up mega-rounds. Meta excluded; it's a public company.
Valuation growth over time
Confirmed anchor points only — ambiguous or conflicting interim marks omitted (see footnotes).
Footnotes & conflicts
- ·OpenAI's Feb 27 / Mar 31, 2026 figures are one round, not two — the announce and close of the same $122B raise.
- ·Anthropic Series G lead investors conflict across sources (GIC/Coatue vs. D.E. Shaw/Dragoneer/Founders Fund/ICONIQ/MGX) — likely co-leads with no single agreed lead.
- ·xAI never officially labeled a "Series D" — the Sept 2025 $10B/$200B round is the closest analogue.
- ·xAI's Series C valuation ($50B vs. $45B) and July 2025 raise valuation ($113B vs. $150B) both have unresolved cross-source conflicts.
- ·Employee counts for all three labs are estimate-only (no official disclosures): OpenAI ~4,500–7,850; Anthropic ~2,300–5,189; xAI ~4,000–5,479.
- ·Revenue run-rate figures (OpenAI ~$20B ARR, Anthropic $47B, xAI ~$500M AI-specific) are third-party/media estimates pending each company's still-confidential S-1.
- ·All three labs have confidentially filed IPO paperwork as of mid-2026; xAI's path went through the SpaceX merger and IPO'd June 12, 2026 at ~$1.77T combined valuation.
The week the frontier moved
Zooming back in — funding rounds and launches from the last 30 days in order. Click any marker for detail.
Reception & controversy
First paid API model from Meta — press framed it as “chasing Anthropic and OpenAI” (CNBC) and “Meta Starts Charging for AI” (Bloomberg). Reviewed as credible but not frontier-leading, and priced below every rival.
Cursor's CEO called it his team's daily driver; best agentic tool-use score on Artificial Analysis's board. Community reaction centered on Musk political-bias trust concerns — the loudest theme on Hacker News — plus an open EU DSA investigation.
Released only after a 12-day, government-brokered CAISI safety review — the standout story. Driven by Sol's top bio-risk score (68.4% on a pathogen capability test) and near-classified cyber capability; OpenAI publicly pushed back on the precedent.
US Commerce Dept forced a global suspension June 12 after a jailbreak reportedly leaked Mythos-tier capability through Fable 5's guardrails. Anthropic argued the same exploit could hit GPT-5.5, unrestricted. Restored July 1; two-thirds of enterprise users had already hedged with fallback models.
Reading the data
- ·GPT-5.6 discloses no SWE-bench, GPQA, AIME, or ARC-AGI scores at launch — a genuine gap in OpenAI's own materials, not a research miss.
- ·Claude Fable 5's headline benchmark figures (95% SWE-bench Verified, 80.3% SWE-bench Pro) come from third-party aggregators, not Anthropic's launch post, which speaks mostly in qualitative superlatives.
- ·Grok 4.5's “1.5T-parameter V9” architecture claim is leak-sourced (CometAPI) and unconfirmed by xAI directly.
- ·Muse Spark 1.1 has almost no independently disclosed quantitative benchmarks — Meta's post is largely qualitative marketing copy.
- ·Terminal-Bench 2.1 scores for Fable 5 conflict across sources (88.0% vs. 83.4%) — likely different eval harnesses; both are noted.
- ·Meta and xAI cumulative funding totals are third-party (Tracxn) rollups, not company-published lifetime figures.
- ·Meta is a public company — its “funding” isn't comparable to the VC rounds raised by the other three labs.