1Executive Summary

April 2026 produced a dense sequence of frontier model releases. Anthropic shipped Claude Opus 4.7 on the 16th and confirmed but withheld Claude Mythos on the 7th, citing internal AI Safety Levels protocols (ASL-4) triggered for the first time in commercial AI history. OpenAI released GPT-5.5 on the 23rd, six weeks after GPT-5.4. Google released the Gemma 4 family under Apache 2.0 on the 2nd. Meta shipped Llama 4 Scout and Maverick on the 5th and reversed its open-weights posture three days later with the proprietary Muse Spark. The closing week added Moonshot AI Kimi K2.6, DeepSeek V4 at 1.6 trillion parameters, and Grok 4.3 from xAI within days of each other.

Coding capability was the focal axis of competition. Opus 4.7 set new highs on SWE-bench Verified (87.6%) and SWE-bench Pro (64.3%); GPT-5.5 led Terminal-Bench 2.0 (82.7%) and OSWorld-Verified (78.7%). Reasoning benchmarks such as GPQA Diamond have effectively saturated at the frontier — Opus 4.7, GPT-5.5, and Gemini 3.1 Pro all sit within 0.4 points of each other.

Industry context was dominated by the closing of Q1 2026 venture data: $297 billion deployed globally, with AI-related companies absorbing 81%. SpaceX completed the acquisition of xAI on terms valuing the combined entity at $1.25 trillion. The European Commission moved closer to enforcement of the EU AI Act, with the Digital Omnibus on AI still under negotiation between Parliament and Council. Compute infrastructure constraints — measured in gigawatts of grid demand rather than dollars — became a structural concern that frontier labs and hyperscalers addressed openly for the first time.

2Large Language Models

Eight frontier or near-frontier models reached general availability or limited preview in April. The cadence is a continuation of the pattern established in Q1, in which release windows shortened to roughly six weeks at OpenAI and roughly two months at Anthropic, with Chinese labs and Meta releasing on overlapping but independent schedules.

2.1Releases and updates

Google released the Gemma 4 family on April 2 under the Apache 2.0 license. The family contains four variants ranging from 2.3 to 31 billion parameters, all natively multimodal across text, image, and video; the two larger variants additionally support audio. The 31B Dense variant ranks third globally on Arena AI among open models. Distribution from day one included Hugging Face, Ollama, Kaggle, and Google AI Studio.

Meta released Llama 4 Scout and Llama 4 Maverick on April 5. Both are open-weight Mixture of Experts models trained with native multimodality from pretraining rather than via later adapter integration. Scout offers a 10-million-token context window; Maverick a one-million-token window. Three days later, on April 8, Meta released Muse Spark — the company's first proprietary model. Muse Spark scores 52 on the Artificial Analysis Intelligence Index. The proprietary release marked an explicit reversal of the open-weights posture that had defined Meta's AI strategy through the Llama family.

Anthropic confirmed the existence of Claude Mythos on April 7 under what the company calls Project Glasswing — an access programme limited to roughly fifty enterprise partners. Mythos scored 93.9% on SWE-bench Verified and 94.6% on GPQA Diamond, both above any publicly released model. Anthropic chose not to release Mythos to general availability, citing the model's autonomous identification of thousands of zero-day vulnerabilities during testing as triggering AI Safety Level 4 protocols. The company stated the model would not be made generally available in the near term.

On April 16 Anthropic released Claude Opus 4.7 to general availability across Claude products, the API (claude-opus-4-7), Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot Pro+, Business, and Enterprise tiers. The release continued Anthropic's roughly two-month cadence following Opus 4.6 in February.

OpenAI released GPT-5.5 on April 23, internal codename Spud. The release was OpenAI’s first fully retrained base model since GPT-4.5 and is natively omnimodal — text, image, audio, and video processed through a unified architecture rather than stitched encoders. API access followed one day later on April 24. GPT-5.5 ships in three consumer surfaces: default GPT-5.5, GPT-5.5 Thinking, and GPT-5.5 Pro.

The closing week produced near-simultaneous releases from Moonshot AI (Kimi K2.6), DeepSeek (V4 at 1.6 trillion parameters, the largest publicly announced model), and xAI (Grok 4.3, promoted out of preview). Alibaba shipped Qwen 3.6-Plus, Zhipu AI released GLM-5.1 under MIT license, and Arcee released Trinity at 400 billion parameters.

2.2Benchmarks and capabilities

Coding capability was the principal axis of differentiation in April. Claude Opus 4.7 scored 87.6% on SWE-bench Verified, up from 80.8% on Opus 4.6 and ahead of GPT-5.4 at 82.0% and Gemini 3.1 Pro at 80.6%. On the harder SWE-bench Pro, Opus 4.7 reached 64.3%, leading GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%). On CursorBench, an evaluation produced by the Cursor IDE on real developer workflows, Opus 4.7 jumped from 58% to 70%.

GPT-5.5 led different coding-adjacent benchmarks. Terminal-Bench 2.0, which evaluates command-line task execution, recorded 82.7% for GPT-5.5 against 69.4% for Opus 4.7 and 68.5% for Gemini 3.1 Pro. OSWorld-Verified, which measures autonomous desktop task completion, showed GPT-5.5 at 78.7% and Opus 4.7 at 78.0%; both pass the 72.4% human expert baseline.

Reasoning benchmarks showed convergence. On GPQA Diamond, Opus 4.7 scored 94.2%, GPT-5.5 scored within the same range, and Gemini 3.1 Pro Preview reached 94.3%. Mythos in restricted access leads at 94.6%. The narrowness of these gaps — within a single percentage point at the top — has effectively saturated the benchmark for the publicly available frontier.

Long context capability progressed. Gemini 3.1 Ultra entered preview in April with a 2-million-token context window operating natively across text, image, audio, and video. Claude Opus 4.7 and GPT-5.5 each retain a one-million-token input context. Llama 4 Scout reached 10 million tokens — the largest published context window for an open-weight model.

Mythos's withholding from general release introduces a new pattern. Claude Mythos is the first model from a major lab to be completed and then withheld on safety grounds. The decision turned on the model's demonstrated ability to identify zero-day vulnerabilities autonomously during red-team testing, which Anthropic determined exceeded the controls applicable to general public access. Whether other labs adopt similar self-imposed restraint remains to be seen.

2.3Pricing and access

Claude Opus 4.7 holds Opus 4.6's pricing of $5 per million input tokens and $25 per million output tokens. The new tokenizer encodes the same text into 1.0 to 1.35 times more tokens than Opus 4.6, so equivalent prompts incur a higher dollar cost despite identical per-token rates. The model is available across Anthropic's full distribution: claude.ai (Pro, Max, Team, Enterprise), the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and Snowflake Cortex AI. GitHub Copilot rolled out Opus 4.7 with a 7.5x premium multiplier as part of promotional pricing valid until April 30.

GPT-5.5's API pricing doubled relative to GPT-5.4: $5 per million input tokens and $30 per million output tokens, against $2.50 and $15 previously. GPT-5.5 Pro held at $30 and $180, unchanged from GPT-5.4 Pro. OpenAI states that GPT-5.5 uses approximately 40% fewer output tokens than GPT-5.4 on equivalent Codex tasks, partly offsetting the per-token increase. Batch and Flex pricing is half the standard rate; Priority pricing is 2.5x. On April 9 OpenAI also introduced a $100 ChatGPT Pro tier between Plus and the existing $200 Pro, the first subdivision of the Pro tier.

Gemini 3.1 Pro retains its undercut position at approximately $2 per million input and $12 per million output tokens, alongside its 2-million-token preview context. Open-weight options expanded materially in April — Gemma 4, Llama 4, GLM-5.1, Qwen 3.6-Plus, and DeepSeek V4 are downloadable for self-hosting at no licence fee. Arena AI and other independent leaderboards now place several open models within ten percentage points of the closed frontier on standard benchmarks.

3Generative Media

April activity in generative media was concentrated in distribution and integration rather than in flagship-model releases. The pattern of recent quarters — model capability advancing fastest in editor-integrated workflows rather than in standalone text-to-image, text-to-video, or text-to-music interfaces — continued.

3.1Image

OpenAI launched ChatGPT Images 2.0 on April 21, replacing GPT Image 1.5. The new system supports a Thinking Mode that integrates reasoning with web search and adds consistency across up to eight images per prompt. Distribution covers ChatGPT Plus, Pro, Business, and Enterprise; free-tier users remain on the previous version.

Black Forest Labs, Stability AI, and Midjourney did not release new flagship image models in April. The image generation segment has shifted toward integration with editor environments rather than standalone model launches; image work increasingly enters workflows through ChatGPT, Google Vids, and Adobe Firefly rather than through dedicated text-to-image interfaces.

3.2Video

Google Vids gained native video generation through Veo 3.1 on April 2, available to anyone with a Google account at the rate of ten free generations per month. Google AI Pro and Ultra subscribers receive higher quotas, with up to one thousand Veo generations monthly available to Ultra accounts. The same release added a Chrome extension for screen recording and direct YouTube publishing.

Runway, Pika Labs, Kling AI, and Luma Labs continued iterative product updates in April without flagship-level releases. The market for AI video generation has consolidated around three high-end systems — OpenAI Sora 2.0, Runway Gen-4, and Veo 3.1 — with Pika and Kling AI competing on accessibility and music-aware workflows.

3.3Music and audio

Google released Lyria 3 and Lyria 3 Pro inside Google Vids on April 2 as the music-generation counterpart to the Veo release. Custom music generation is gated to Google AI Pro and Ultra subscribers. Suno and Udio did not release new flagship models in April. Udio remained constrained by the October 2025 Universal Music Group settlement, which limits paid users to streaming-only access without download capability. ElevenLabs continued integration partnerships, including the LTX Studio audio-to-video pipeline launched in January.

4Coding and Developer Tools

The April releases of Claude Opus 4.7 and GPT-5.5 were both positioned by their developers around coding and agentic workflows rather than chat or general reasoning. The framing reflects where revenue concentration has gone: Cursor at $2 billion annualized revenue by early 2026, Claude Code at $2.5 billion annualized run rate, and GitHub Copilot approaching $1 billion.

4.1Releases

Claude Opus 4.7 became the default coding model in Claude Code on April 16. The release added an xhigh effort level between high and max, giving finer reasoning-depth control without the full latency cost of max. The model also introduced explicit task budgets in beta and a /ultrareview command in Claude Code. Adaptive thinking replaced extended thinking budgets — setting budget_tokens on the API now returns a 400 error.

GPT-5.5 powered Codex from April 23 with a 400,000-token context window for paid plans. OpenAI granted Pro users 2x Codex usage allowance through May 31. The model is positioned explicitly as an agent runtime: the company’s announcement framed it as suitable for "messy, multi-part" tasks that require planning, tool use, self-checking, and persistence rather than single-turn chat completion.

Cognition completed the acquisition of Windsurf for $250 million during April, following Google’s acqui-hire of the Windsurf founders for $2.4 billion in July 2025. The transaction integrates Windsurf’s IDE and Cascade agent into Cognition’s Devin product line. Sourcegraph spun out Amp as a separate company.

4.2Pricing and licensing

Cursor maintained its $20 per month Pro tier. Devin pricing remained at $20 Core plus $2.25 per Agent Compute Unit, the structure introduced in late 2025 to broaden access from the original $500 monthly enterprise tier. GitHub Copilot rolled out Opus 4.7 with a 7.5x premium multiplier through April 30 as promotional access; the post-promotion multiplier had not been published at the close of the month. GPT-5.5 rolled out across Copilot Pro+, Business, and Enterprise alongside the OpenAI Codex integration.

4.3Adoption signals

JetBrains published the second wave of its AI Pulse developer survey, with January 2026 data drawn from more than ten thousand professional developers. Among specialised AI coding tools, GitHub Copilot retained the highest awareness and adoption — 76% awareness, 29% adoption at work — but its growth has stalled since 2025. Cursor sits at 69% awareness with adoption growth slowed. Claude Code grew from 3% adoption at work in mid-2025 to 18% in January 2026, reaching 24% in the United States and Canada, with the highest customer satisfaction (91%) and net promoter score (54 on a -100 to +100 scale) in the surveyed cohort. The same survey reported that 74% of professional developers worldwide had adopted at least one specialised AI coding tool by January 2026.

5Industry Trends

Q1 2026 venture and exit data closed during April. Capital concentration in AI reached levels without precedent in the venture record, and the structural constraints on continued AI deployment — power, custom silicon, regulatory readiness — came into clearer view as the EU AI Act enforcement deadline approached.

5.1Funding and valuations

Q1 2026 closed with the highest quarterly venture capital investment ever recorded. KPMG measured $330.9 billion globally; Crunchbase logged $297 billion concentrated in approximately 6,000 companies; CB Insights reported $285.5 billion. The variance reflects different methodologies and reporting cutoffs. AI-related companies absorbed approximately 81% of total deployment.

Four megarounds dominated the quarter: OpenAI raised $122 billion at a valuation that placed the company higher than all but a handful of large-cap public technology companies; Anthropic raised $30 billion in a Series G led by GIC and Coatue at a $380 billion post-money valuation; xAI raised $20 billion before its acquisition by SpaceX; and Waymo raised $16 billion. Databricks closed a $7 billion round; Polymarket $2.6 billion; Shield AI $2.3 billion. The four leading rounds collectively totalled $188 billion, exceeding all of 2024's global venture deployment.

Two AI foundation labs based in mainland China entered public markets in Q1 via Hong Kong: Z.ai (Zhipu AI) and MiniMax, each valued above $6 billion at listing. The Hong Kong Stock Exchange has emerged as a viable IPO venue for Chinese AI companies absent direct US listings. Anthropic was reported in late April to be running at a $30 billion annualized revenue rate, with secondary investor offers reportedly arriving at valuations approaching $800 billion and early IPO discussions under way. OpenAI exceeded $25 billion in annualized revenue.

5.2Regulation and legal

The European Commission issued €63.2 million in support of AI innovation in health and online safety on April 21. The 2 August 2026 enforcement date for the bulk of EU AI Act obligations remained in force at the close of the month, although the Digital Omnibus on AI — adopted by the Commission in November 2025 — continued to be negotiated by the Parliament and Council. The Digital Omnibus proposes a maximum postponement to 2 December 2027 for certain high-risk system obligations under Annex III. Article 50 transparency requirements for AI-generated content remained on the original 2 August 2026 schedule under all current proposals.

State-level regulation in the United States continued to fill the gap left by the absence of a federal framework. The Texas Responsible Artificial Intelligence Governance Act took effect on January 1, 2026; the Utah Artificial Intelligence Policy Act applies disclosure requirements to deployers of generative AI in regulated and consumer transactions; the Colorado AI Act becomes applicable in June 2026. New York City and the federal Equal Employment Opportunity Commission moved on enforcement actions related to AI-driven hiring tools.

The principal generative AI copyright cases — New York Times v. OpenAI in the Southern District of New York, and Getty Images v. Stability AI in the United Kingdom and the United States — entered decisive phases in April. Court rulings on whether training on copyrighted data constitutes fair use are expected during Q2 2026.

5.3Mergers, acquisitions, exits

SpaceX completed the acquisition of xAI on terms that valued the combined entity at $1.25 trillion and the xAI business itself at $250 billion. The transaction is the largest M&A in corporate history by absolute value at signing. The combined entity owns Tesla's autonomous-driving programmes, the Grok model family, the X social platform, and the Starlink satellite network.

OpenAI continued an active acquisition programme; the company completed six acquisitions during 2026 by mid-April, equalling its full-year acquisition count for 2025. Anthropic completed one publicly disclosed acquisition in 2026 — Vercept, a software development startup founded in 2024 — adding to the 2025 acquisitions of Humanloop and Bun.

Total venture-backed M&A in Q1 reached $56.6 billion per Crunchbase, the third-highest quarterly total since the 2022 downturn. Total Q1 exit value crossed $413.5 billion globally per KPMG, the highest level since Q4 2021.

5.4Infrastructure

Compute infrastructure constraints emerged as a binding capacity for AI deployment. Projections for the United States indicate a 9 to 18 gigawatt electricity shortfall by 2027 driven by AI data centre construction. Microsoft’s announced restart of the Three Mile Island nuclear facility, originally signalled in 2024, advanced through April. Amazon continued exploration of dedicated nuclear sites for AWS data centre power.

Custom silicon partnerships diversified during the quarter. Google extended its TPU partnership with Broadcom; Anthropic signed a multi-year compute commitment with CoreWeave; Intel announced a collaboration with Google on next-generation AI accelerators. AMD continued GPU shipments at scale and increased its participation in AI venture rounds via AMD Ventures, including the Cohere round and the World Labs financing.

A research result attributed to a method called TurboQuant — applying 3-bit quantization to KV cache memory without measurable accuracy loss — was reported during the quarter, with claims of up to eight-fold speedup in attention computation and at least six-fold reduction in memory usage. Independent verification of the published method was pending at the close of April. Arista Networks raised its 2026 revenue outlook to $11.25 billion, citing AI cluster deployment as the principal driver.

6Monthly Recap

Chronological summary of the month’s reportable events. Sources are linked in the corresponding section above.

7Outlook

GPT-6 timing remains uncertain. Prediction markets had assigned April 2026 some probability of release; that has now slipped. Sam Altman’s most recent on-record statement was "a few more weeks." Whether the next OpenAI release continues the GPT-5.x line or marks a generational increment to GPT-6 has not been resolved.

xAI has indicated Grok 5 for the second quarter, with reported architectural details pointing to a 6-trillion-parameter Mixture of Experts. The validity of these reports is partial; treat the parameter count as unconfirmed.

DeepSeek's R2 successor has been reported as delayed by performance shortfalls and chip-export-control constraints affecting access to high-end NVIDIA hardware. A Q2 release window appears unlikely on current information.

Claude Opus 4.7 may itself yield to a successor before September. Anthropic's roughly two-month cadence — Opus 4.5 in November 2025, Opus 4.6 in February 2026, Opus 4.7 in April 2026 — implies an Opus 4.8 release in the early summer, with possible introduction of a Mythos-class model in restricted general availability if internal safety reviews complete.

The Digital Omnibus on AI is expected to reach political agreement in the Council before June. Whether the proposed postponement of certain high-risk obligations to 2027 takes legal effect before the original 2 August deadline will determine the compliance posture for several thousand companies operating in or selling into the European market.

The first rulings in the principal generative-AI copyright cases — New York Times v. OpenAI and Getty Images v. Stability AI — are expected during Q2.