May 2026's release wave: every launch, priced and ranked.

May 2026 delivered more frontier-model launches in a single month than any prior period in 2026 — Gemini 3.5 Flash, Composer 2.5, Grok Build 0.1, Gemini Omni Flash, Antigravity 2.0, Managed Agents in the Gemini API, Anthropic self-hosted sandboxes, MCP tunnels, Microsoft Copilot Studio computer-use GA, and GLM-5.1 — all within 22 calendar days. This tracker assembles every confirmed launch with verified pricing, benchmark anchors, and availability data, sourced directly from vendor documentation retrieved on May 24, 2026.

The stakes behind the velocity: three separate pricing promos expire within eight days of publication. GitHub Copilot transitions from request-based to AI-credit billing on June 1 — and per-credit dollar pricing has not been published. Gemini 3.5 Pro is committed for June 2026 per Sundar Pichai's I/O stage remarks. The decisions teams make about model selection this week are pricing decisions, not capability decisions — and the promotional-rate window is closing.

This post covers: the master May 2026 model-launch tracker (Table 1), the 22-day pricing-event chronology (Table 2), the cost-per-task comparison matrix (Table 3), benchmark normalization rebased to Opus 4.7 = 100 (Table 4), the intro-promo cliff pattern as a recurring industry dynamic, the enterprise-deployment wave that completed the Big Four AI-deployment matrix, and the June outlook. For the Gemini 3.5 Flash deep-dive, see our Gemini 3.5 Flash benchmarks and API guide. For the comparable prior-month format, see our April 2026 coding-assistants tracker.

The table below is the canonical May 2026 model-release record. All pricing figures are live-verified from vendor documentation as of 2026-05-24 — specifically ai.google.dev/gemini-api/docs/pricing for Gemini rates and docs.x.ai/docs/models for Grok rates. Benchmark anchors marked ⚠️ are vendor-controlled or vendor-published — cross-variant comparison with SWE-Bench Verified is not appropriate for these rows. Rows for April 2026 launches are included as context anchors. For the full April 2026 launch record, see our March 2026 release-wave tracker and Q2 2026 release-velocity index.

April 2026 context anchors (included for tracker completeness): Claude Opus 4.7 (April 16, $5/$25 per Mtok, 87.6% SWE-Bench Verified), Grok 4.3 (April 17, $1.25/$2.50 per Mtok, 1M context), GPT-5.5 (April 23, $5/$30 per Mtok with 272K long-context surcharge), DeepSeek V4 Preview (April 24, 1M context, preview status). For the Opus 4.7 deep-dive, see our Claude Opus 4.7 complete guide and the Opus 4.7 1M-context cost strategy guide. For the GPT-5.5 surcharge mechanics, see our GPT-5.5 complete guide.

Seven separate pricing events and ten-plus model launches compressed into 22 calendar days. The table below is the complete record from the April 30 Opus 4.7 Copilot promo expiry through the June 1 GitHub Copilot billing transition — every date a pricing or availability signal changed. No competitor publication has assembled the May 2026 pricing-event chronology in one place; this becomes the evergreen reference for retrospectives and Q2 2026 wrap-ups.

For further context on the Copilot billing transition and premium-request multiplier mechanics, see our Copilot vs Cursor vs Windsurf comparison. For the broader agentic-coding platform matrix across 20 tools, see the Q2 2026 20-platform agentic-coding matrix.

Every prior major Google I/O shipped a Pro model first. Gemini 1.5 Pro appeared ahead of 1.5 Flash. Gemini 2.0 Pro headlined before Flash variants followed. The pattern was so consistent that developer convention assigned “Pro” as the keynote anchor and “Flash” as the follow-on efficiency tier.

I/O 2026 inverted this. Gemini 3.5 Flash reached GA on May 19 — keynote day. Gemini 3.5 Pro was held for June 2026 with Sundar Pichai's stage commitment: “Already being used internally and we look forward to rolling it out next month.” (Sundar Pichai, I/O 2026 keynote, May 19, 2026.)

Two explanations, not mutually exclusive: (1) Compute scarcity — Pro models require significantly more inference compute per token; launching Flash first may reflect that Google can serve Flash at scale while Pro demand is still being ramped at the infrastructure layer. (2) Agent-era positioning— Gemini 3.5 Flash, at $1.50 input and "less than half the cost of other frontier models" per Google's I/O 2026 100 Things recap, is explicitly positioned as the agent-optimized tier. Agentic workloads run millions of API calls per task; the economics only work at Flash pricing. Google may be signaling that Flash is the correct model for the agent era — and Pro is the fallback for tasks that genuinely require maximum capability.

The interpretation matters for practitioners. If Google is correct that Flash-class models are the right default for agent loops, then Composer 2.5 ($0.50/$2.50), Grok Build ($1.00/$2.00), and Gemini 3.5 Flash ($1.50/$9.00) represent the three most cost-relevant pricing points for 2026 agentic deployment — not Opus 4.7 or GPT-5.5. For the head-to-head analysis of these three on agentic coding tasks, see our Gemini 3.5 Flash vs GPT-5.5 vs Opus 4.7 head-to-head.

Pichai's use of Antigravity to co-develop Flash itself is the strongest internal signal: Antigravity 2.0 was used to build the model that now powers Antigravity. The recursive loop — Flash accelerating Flash — is both a product story and an architecture thesis about where the performance ceiling actually sits in this generation of models.

Most model-pricing comparisons print headline $/Mtok rates and stop there. Four levers — tokenizer overhead, long-context surcharges, cache hit rate, and Fast-mode pricing — routinely produce 2-10× effective-cost differences that never appear in the headline rate. The matrix below makes each lever explicit. Per-task math: $/Mtok-in × 0.1 (100K input tokens) + $/Mtok-out × 0.02 (20K output tokens), uncached. All rates from vendor docs retrieved 2026-05-24.

The Opus 4.7 tokenizer overhead is a separate risk category from the long-context surcharge. Anthropic's pricing documentation discloses: “Opus 4.7 uses a new tokenizer compared to previous models. This new tokenizer may use up to 35% more tokens for the same fixed text.” For workloads migrating from Sonnet 4.6 to Opus 4.7, this overhead means effective per-task cost can reach $1.35 even at the $5/$25 headline rate — a 35% cost premium invisible in the API pricing page. For the full cost-strategy analysis at 1M context scale, see our Opus 4.7 cost strategy guide.

Raw SWE-Bench Verified percentages are not directly comparable across vendors because different evaluation harnesses, task selections, and instruction formats produce materially different scores on the same underlying model. Rebasing to a common anchor (Opus 4.7 = 100) makes the relative capability difference readable: a score of 80 means that model completes roughly 80% of the tasks Opus 4.7 would complete in the same harness, all else equal.

Critical caveat: Composer 2.5 and Gemini 3.5 Flash do not publish SWE-Bench Verified scores at launch. Both publish only vendor-controlled or vendor-published benchmarks (CursorBench v3.1, SWE-Bench Multilingual, Terminal-Bench 2.1, MCP Atlas). Cross-variant comparison with SWE-Bench Verified for these models is not appropriate per Day 02 SWE-Bench methodology research. For the full leaderboard analysis, see our Q2 2026 SWE-Bench leaderboard analysis.

The “open-weight base + proprietary RL” architecture that Composer 2.5 demonstrates — Moonshot's open-weight Kimi K2.5 as the foundation, with ~85% of total compute in Cursor's proprietary RL post-training — is the clearest commercial example of a two-layer split in 2026. The controlled experiment is unusually clean: Composer 2 and Composer 2.5 share the same base model (K2.5), separated by 60 days and 25× more synthetic tasks during RL post-training. The result — +11.0 points on CursorBench v3.1, +6.1 on SWE-Bench Multilingual, +7.6 on Terminal-Bench 2.0 — is the strongest single data point in 2026 for “post-training is the new scaling axis.” For the Chinese open-weight context for K2.5 and Kimi K2.6, see our Kimi K2.5 agent swarm guide and the Q2 2026 Chinese AI market share report.

Four separate intro discounts are active across the May 2026 frontier model landscape. Three expire within eight days of this publication date. This is not a coincidence — it is a pattern that has now repeated across every major launch since Opus 4.7 in April.

The four active promos (as of May 24, 2026):

1. Composer 2.5 first-week 2× rate. Standard: $0.25/$1.25 during promo period (effectively half the list price). Promo ends May 25 — tomorrow from publication. Standard rate of $0.50/$2.50 applies from May 26. Source: Cursor blog — Composer 2.5 launch post, May 18, 2026.

2. Codex Pro 2× promo.“Double your normal Codex usage on the $100/month tier until May 31, 2026.” Seven days from publication. Source: OpenAI Codex pricing page, retrieved 2026-05-24.

3. SuperGrok Heavy $99/mo intro.Six-month introductory rate, then $300/mo list price. Grok Build access requires SuperGrok Heavy — meaning the effective cost of Grok Build roughly triples after the intro period ends. DISAMBIGUATION: Standard SuperGrok at $300/year (~$25/mo) is a different tier. Heavy is $300/month post-promo. Both display “$300” in different billing contexts.

4. Opus 4.7 Copilot promo — already expired April 30.The 7.5× premium-request multiplier on GitHub Copilot doubled to 15× on April 30. For context: Copilot Pro includes 300 premium requests. At 15×, that is effectively 20 Opus prompts per billing period — a meaningful effective-cost change for teams that built workflows at the 7.5× rate.

The strategic implication: introductory rates are now the industry default for frontier model launches. Every model in the May cohort launched with a time-limited discount. Teams that evaluate models during the promo window and budget accordingly will face a 2–3× cost increase within 90 days of every major launch. The right organizational response is to build workflows that are cost-portable across model tiers — not to optimize exclusively for the intro price. For the efficient frontier of price vs performance across the full Q2 2026 model landscape, see our Q2 2026 price-vs-performance efficient frontier.

The May 4–21 deployment wave did not just deliver new models — it completed the Big Four audit-firm AI-deployment matrix, the clearest signal that enterprise AI has moved from pilot to volume contract infrastructure.

The sequence: Anthropic + Blackstone $1.5B JV (May 4) → OpenAI Deployment Co $4B (May 11) → SAP + Claude Joule (May 12) → PwC + Anthropic 30K seats (May 14) → KPMG + Anthropic 276K seats (May 19) → EY × Microsoft $1B direct (May 21). With Deloitte's Microsoft-adjacent posture, three of the four Big Four firms are now on Claude at scale. The EY + Microsoft $1B deal closes the matrix — the four largest professional services firms in the world have each made an explicit, public AI deployment commitment at the volume level.

The parallel Anthropic counter-programming at Code with Claude London (May 19–20) — shipping self-hosted sandboxes and MCP tunnels the same day as I/O — was not coincidental timing. The combination of the KPMG deal announcement (276K seats) and the enterprise sandbox features signals Anthropic's Q2 2026 go-to-market: land large-volume enterprise contracts, then deepen the moat with infrastructure that only runs in the customer's VPC.

For practitioners running AI strategy at large organizations: the Big Four adoption pattern is the clearest proxy signal for where enterprise AI spend is moving in H2 2026. Three-of-four on Claude is not coincidence — it reflects a specific set of security, compliance, and contract terms that Anthropic has built for the professional-services buyer. Our AI transformation advisory services work with organizations navigating these same vendor-selection and deployment-security decisions.

June 2026 carries three confirmed hard dates and one major strategic unknown that will shape the next quarter of AI pricing decisions.

June (early) — Gemini 3.5 Pro rollout.Sundar Pichai committed on stage at I/O: “Already being used internally and we look forward to rolling it out next month.” No launch date within June has been specified. Pricing is not published. This will be the most anticipated model release of Q2 2026: if Google holds the Flash-first pattern, 3.5 Pro will land as a premium tier above 3.5 Flash — potentially with a tiered surcharge above a context threshold, following the Gemini 3.1 Pro pattern (200K threshold, 2× rate above). Watch the Gemini API pricing page for the first indication of launch.

June 1 — GitHub Copilot AI-credit billing begins.As of May 24, 2026, per-credit dollar pricing has not been published. Teams running Copilot at scale need to download their April usage reports (available from May 12, 2026) and model the credit-vs-request translation before June 1. The premium-request multipliers remain the best available proxy: Opus 4.7 at 15× and GPT-5.5 at 7.5× define the outer bounds of per-task cost in AI credits. For the detailed premium-request mechanics, see the GitHub Copilot requests documentation.

May 31 — Codex Pro 2× promo expires. Seven days from publication. Teams on the $100/mo Codex Pro plan will see their effective Codex capacity halve after the promo window closes.

Strategic unknown: Cursor + SpaceX. The SpaceX $60B acquisition option (April 21–22, 2026) triggers approximately 30 days after the SpaceX IPO, currently targeted for June 12, 2026, at a $1.75T valuation. If the IPO proceeds and the option exercises, the acquisition closes approximately July 2026. Cursor's training roadmap (Composer 3 on Colossus compute) and pricing strategy would become subject to SpaceX + xAI decisions rather than Cursor's independent roadmap. This is the largest single structural uncertainty in the agentic-coding pricing landscape for H2 2026. For forward-looking analysis, see our 30 agentic AI predictions for H2 2026.

For the full Q2 2026 model-launch arc across March, April, and May, see our Q2 2026 release-velocity index. For the H1 2026 retrospective placing May 2026 in context, see the H1 2026 agentic-coding retrospective.