Google Names the Agentic Era at I/O as Anthropic Nears $900 Billion
Gemini Spark launched as a 24/7 background agent yesterday at Google I/O 2026, while Anthropic is in final talks for a $50B raise that would push its valuation above OpenAI's.
The story
On Tuesday in Mountain View, Sundar Pichai stepped on stage at Google I/O 2026 and declared the agentic era open for business. The headline product was Gemini Spark: a personal AI agent running 24/7 on dedicated Google Cloud VMs, executing tasks in the background even when your laptop is closed. Alongside it came Gemini 3.5 Flash — available now in the API at $1.50/$9 per million tokens — and Gemini Omni, a world model that generates video and audio from any input. Google also reported the Gemini app hit 900 million monthly active users, up from 400 million a year ago. These announcements arrived as Bloomberg reported Anthropic is in final talks for a $50 billion fundraise at a $900 billion valuation — higher than OpenAI's — led by Dragoneer and Sequoia. The pattern across all three top labs is now unmistakable: the competition has moved from who has the best model to who owns the agent layer. Gemini Spark, Claude Code, and OpenAI's Deployment Company are all placing the same bet from different positions. Gemini 3.5 Pro, due next month, is the next flashpoint.
Who moved this cycle
| Player | What changed | Consequence |
|---|---|---|
| Shipped Gemini 3.5 Flash to the API at $1.50/$9 per 1M tokens, 4x faster than frontier peers · Launched Gemini Spark, a 24/7 personal agent running on dedicated Cloud VMs · Unveiled Gemini Omni, a world model that generates video, image, and audio from any input · Reported Gemini app at 900M monthly active users, up from 400M a year ago | The first credible always-on background agent from a major lab changes the benchmark from 'best single response' to 'longest reliable task chain' — Anthropic's Claude Code and OpenAI's Deployment Company now have a direct competitor at the agent layer | |
| Anthropic | In final talks to raise $50B at a $900B valuation (Bloomberg, May 12), surpassing OpenAI's $852B post-money · Already at $30B ARR run-rate as of April (company-announced) · Splitting subscription into interactive and programmatic credit pools from June 15, triggering developer backlash · Named #1 on CNBC Disruptor 50 on May 19 | If the $900B round closes, Anthropic has runway to match Google's infrastructure spending and undercut OpenAI's enterprise pricing; the June 15 subscription split risks pushing cost-sensitive developers toward DeepSeek V4-Flash and Gemini 3.5 Flash |
| Meta | 8,000-employee layoffs begin today (May 20), announced April 17, alongside cancellation of 6,000 open requisitions · $145B committed to AI infrastructure in 2026 · Muse Spark (closed-source, launched April 8 from Meta Superintelligence Labs led by Alexandr Wang) benchmarks at frontier level | Meta's math: AI capex returns exceed returns on those 14,000 headcount positions — if that calculation proves correct, every large tech company faces the same pressure to make the same trade |
| OpenAI | Deployment Company now live with 19 global partners — Capgemini, Bain & Company, TPG, Brookfield — following May 12 launch · GPT-5.5 API available at $5/$30 per 1M tokens with 1M-token context · Revenue chief Stuart Dresser said enterprise AI adoption is 'at a tipping point' (May 11) | Wrapping GPT-5.5 in a services network with Bain and Capgemini addresses the last objection from Fortune-500 procurement teams: model access alone never closes an enterprise contract |
| xAI | Grok 4.3 became the default API model on May 15 as older slugs were retired · Speech-to-Text and Text-to-Speech APIs now generally available in 25 languages · Voice cloning from a short audio clip added across TTS and Voice Agent APIs | xAI is expanding from text inference toward a multi-modal API stack competing with ElevenLabs and OpenAI Realtime at the voice layer — an area Anthropic has not entered |
| Alibaba | Qwen AI merged into Taobao shopping app, giving Qwen access to 4B Taobao/Tmall products for conversational purchase and logistics · Consumer Qwen interface hit 300M monthly active users · Cloud Intelligence Group AI-related product revenue posted triple-digit growth for the eleventh consecutive quarter | A 300M-MAU AI shopping layer built on Qwen is a data flywheel with no Western equivalent — every order trains intent models that Alibaba alone controls |
New models on the board
| Player | Model & what changed | Pricing |
|---|---|---|
| Gemini 3.5 Flash GA as of May 19 · 4x faster than frontier-class peers · outperforms Gemini 3.1 Pro on coding and agentic benchmarks · built for long-horizon background tasks via Antigravity harness | $1.50/$9.00 per 1M tokens (3x above Gemini 3 Flash but new capability tier) | |
| Anthropic | Claude Opus 4.7 Released April 16 · SWE-bench Pro 64.3% (up from Opus 4.6's 53.4%, biggest single-version jump in 2026) · enhanced vision with higher-resolution image processing · new tokenizer (up to 35% more tokens on fixed text) | $5/$25 per 1M tokens (unchanged from Opus 4.6) |
| OpenAI | GPT-5.5 Released April 24 · agentic task completion across browser, code, and data · 1M-token context · GPT-5.5 Pro variant for highest-complexity workloads | $5/$30 per 1M tokens (standard) · $30/$180 per 1M tokens (Pro) · batch at half rate |
| DeepSeek | V4-Flash Released April 23 · MIT license with downloadable weights · frontier-class benchmark scores · lowest-priced flagship-class model available | $0.14 per 1M input tokens |
Market signals
| Player | Type | Value | Context |
|---|---|---|---|
| Anthropic | Funding | $50B raise at $900B pre-money valuation (Bloomberg, May 12, unconfirmed by Anthropic) | Dragoneer, Greenoaks, Sequoia, and Altimeter are co-leads; Amazon separately committed up to $25B plus 5GW compute · would surpass OpenAI's $852B post-money valuation · reportedly Anthropic's last private round before IPO |
| Anthropic | Revenue | $30B ARR run-rate (company-announced, April 2026) | Up from roughly $1B in late 2024 — 30x in 16 months · 80% from enterprise API consumption · Claude Code contributing ~$2.5B of the total |
| OpenAI | Partnership | OpenAI Deployment Company launched May 12 with 19 founding partners | TPG as lead, Advent, Bain Capital, Brookfield as co-leads · includes Capgemini and Bain & Company as system integrators · OpenAI acquired ~150-person consulting firm Tomoro to staff the venture |
| Anthropic | Partnership | Enterprise AI joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs at $1.5B valuation (announced May 4) | $300M commitment each from Anthropic, Blackstone, and H&F · mirrors OpenAI Deployment Company structure · targets financial services and healthcare verticals |
| Meta | Hiring | 14,000 positions eliminated (8,000 layoffs + 6,000 cancelled requisitions, begins May 20) | Framed by Zuckerberg as converting headcount budget into AI infrastructure · $145B capex committed for 2026 · Alexandr Wang's Meta Superintelligence Labs is the primary destination for new AI headcount |
| Alibaba | Revenue | Triple-digit AI product revenue growth for 11th consecutive quarter (company-announced) | Qwen consumer MAU hit 300M · Taobao integration gives Qwen access to 4B SKUs and order/logistics agents · Cloud Intelligence Group accelerating |
What builders are saying
Michael Truell, CEO of Cursor. Confirmed Opus 4.7 lifted resolution by 13% over Opus 4.6 on Cursor's internal 93-task benchmark, solving four tasks that neither Opus 4.6 nor Sonnet 4.6 could complete — the specific note was that the hardest 5% of coding problems finally moved.
Theo Browne, founder of T3.gg. Warning to developers about Anthropic's June 15 split: anyone using Conductor, Zed, or claude -p in CI scripts has effectively had their usage cut by 25x because the new credit pool doesn't inherit subscription rates — programmatic usage is now priced separately.
Sundar Pichai, Google CEO, I/O 2026 keynote. On Gemini Spark: 'It's your personal AI agent that helps you navigate your digital life, taking action on your behalf and under your direction. It runs on dedicated virtual machines on Google Cloud seamlessly — you don't need to keep your laptop open to make sure it's running.'
Air Street Press, State of AI May 2026. The April-May 2026 window is the most competitive in AI history: four Chinese labs released open-weight coding models inside a 12-day span — Z.ai GLM-5.1, MiniMax M2.7, Moonshot Kimi K2.6, and DeepSeek V4 — none costing more than a third of Claude Opus 4.7.
Threads worth pulling
Gemini Spark on dedicated Cloud VMs → every Spark task runs compute on Google Cloud → Google charges cloud margin on top of model inference → Spark is a Cloud growth vehicle disguised as a product launch Google has the cheapest path to 'agent always on' because it owns the infrastructure. Every background task Spark runs is a billable Cloud unit. The model is effectively a loss leader for the VM and storage stack — a trick Anthropic and OpenAI cannot copy without their own cloud footprint.
Anthropic's June 15 programmatic credit split → CI/CD pipelines and vibe-coding tools lose subsidy → some developers migrate to DeepSeek V4-Flash at $0.14/1M → DeepSeek gains Western developer mindshare → MIT license means those developers build tooling on open weights Anthropic's subscription restructuring is a pricing experiment on its most loyal and loudest users — developers. If even 10% of Claude Code programmatic users switch to DeepSeek to save 95% on token costs, the open-weight ecosystem gets a massive western on-ramp.
Meta abandons open-source with Muse Spark → Chinese labs (DeepSeek V4, Kimi K2.6, GLM-5.1) fill the open-weight vacuum → Western enterprises adopt Chinese open-weight models in air-gapped infra → export control concerns intensify Llama was the reason enterprise buyers considered open-weights safe — it had a credible American brand behind it. With Muse Spark going closed, the natural open-weight alternatives are now Chinese. That is the geopolitical thread regulators in Brussels and Washington haven't publicly addressed yet.
Trump lifts AI chip ban on China → Nvidia restarts H200 manufacturing for Chinese customers → DeepSeek's next training run has better hardware → inference cost advantage over Western frontier narrows but training efficiency advantage widens The ban lift is counterintuitive: better Nvidia chips in China accelerate Chinese model capability, but Chinese labs are already so efficient on training that the marginal hardware gain goes into larger models, not cheaper inference. The real beneficiary may be Nvidia's stock, not Western AI competitiveness.
What we’ll be watching
- Anthropic's $50B round close — term sheets reportedly due within days of May 12 report; a signed deal by end of May would confirm the $900B valuation and likely trigger Anthropic's IPO timeline discussion
- Gemini 3.5 Pro arrival — Google said 'next month' at I/O; given the Flash launch landed May 19, Pro could be June or earlier; it will be the real test of whether Google can beat Opus 4.7 on complex reasoning
- Meta's 8,000-employee layoffs executing today (May 20) — watch for employee posts and Glassdoor signal on which teams were cut and what projects were cancelled as a result
- Anthropic's June 15 subscription split developer reaction — open questions include whether major tooling vendors like Cursor, Zed, and Replit will absorb the cost or pass it through to end users
- OpenAI Deployment Company first announced enterprise deployments — Capgemini and Bain said to have clients lined up for Q2 2026 go-lives; watch for named Fortune-500 customers
Sources: blog.google, techcrunch.com, cnbc.com, bloomberg.com, 9to5google.com, venturebeat.com, openai.com, anthropic.com, the-information.com, llm-stats.com, buildfastwithai.com, analyticsinsight.net, thenextweb.com, fortune.com, press.airstreet.com.