GPT-5.5, DeepSeek V4, Claude Opus 4.7: The month everything changed again

April 2026 was — and I don’t say this lightly — the densest eight days in the history of AI model releases. Claude Opus 4.7 on April 16. GPT-5.5 on April 23. DeepSeek V4 less than 24 hours later. Then Gemini 3.1 Pro, Llama 4, Qwen 3, and Gemma 4, all in the same six-week window.

I’ve been tracking AI tool releases professionally for two years. This is the one where I had to stop, recalibrate my model routing config, and write a post I’d been putting off: the era of loyalty to a single model is over. The economics don’t allow it anymore.

This post covers every major release from mid-April through mid-May 2026 — not just the models, but the tools built on top of them: coding agents, research platforms, design AI, enterprise copilots. I’ll give you benchmarks where they’re reliable, flag where they’re marketing, and tell you which tools I’d actually bet on for different use cases.

My sample skews toward B2B and developer workflows — US and EU markets primarily. If you’re running consumer apps or operate in regulated industries, your mileage will differ and I’ll say so explicitly.

The three releases that rewrote the leaderboard

Every few months someone writes “the AI landscape is moving fast” and it’s mostly filler. This time it’s actually true in a measurable sense. The SWE-bench Pro ranking flipped four times in eight days.

Claude Opus 4.7 Established

Anthropic’s April 16 release came with a 12-point gain on CursorBench over 4.6 and pushed SWE-bench Verified to 87.6% — the highest any model has posted on that benchmark. For context, GPT-5.5 sits at 58.6% on SWE-bench Pro (a harder variant). Claude Code, Anthropic’s CLI agent powered by Opus 4.7, runs at 91% customer satisfaction and an NPS of 54 per JetBrains’ January developer survey.

The upgrade that actually matters for production teams is task budgets: a configurable ceiling on how many steps an agentic run can take before pausing for human review. Before this, a misconfigured agent could silently burn through quota. I’ve seen this happen at scale — once with roughly $4,000 in unexpected API charges from a loop that ran overnight. Task budgets prevent that class of failure.

One thing to audit before you migrate: Opus 4.7 uses a new tokenizer that produces up to 35% more tokens for identical input text. The rate card is unchanged from 4.6 ($5 input / $25 output per million tokens), but real costs per request can rise. If you’re running automated pipelines, measure actual token consumption before assuming costs stay flat.

GPT-5.5 “Spud” Established

OpenAI launched GPT-5.5 on April 23, 2026 — the first fully retrained base model since GPT-4.5. The codename “Spud” made the rounds after a developer found it in a Codex system prompt. Less amusingly, the same prompt contained an instruction not to mention goblins — a quirk that emerged from the “Nerdy” personality training signal leaking into unrelated outputs. OpenAI has since retired that personality and addressed the issue.

What the model actually delivers: 82.7% on Terminal-Bench 2.0, 78.7% on OSWorld-Verified (simulated desktop tasks, human baseline 72.4%), and 84.4% on BrowseComp. It’s notably stronger than GPT-5.4 on long-horizon agentic tasks — Cursor’s CEO described it as “noticeably more persistent.” On May 5, the lighter GPT-5.5 Instant variant replaced GPT-5.3 as the default for free-tier ChatGPT users. The Pro tier ($200/month) offers the full 5.5 version, with GPT-5.5 Thinking available for complex reasoning workloads.

DeepSeek V4 Established

This is the one MIT Technology Review called “the most significant release since R1.” Released April 24 — less than 24 hours after GPT-5.5 — and that timing felt deliberate.

V4 ships in two variants: V4-Pro (1.6 trillion total parameters, 49B active per token, $1.74/M input) and V4-Flash (284B total, 13B active, $0.14/M input). Both are MIT-licensed, both support 1 million token context windows, and both were trained entirely on Huawei Ascend 910B chips — zero Nvidia GPUs. Whatever you think about US export controls on AI chips, their practical effect on DeepSeek’s trajectory has been approximately zero.

The architectural innovation is Hybrid Attention — combining Compressed Sparse Attention and Heavily Compressed Attention to handle 1M-token contexts at 27% of the FLOPs V3.2 required, and 10% of the KV cache. For a production application running 100 million tokens per month, the difference between Claude Opus 4.7 ($2,500/month) and DeepSeek V4-Flash ($14/month) is real enough that any team not at least testing V4 is leaving money on the table.

The honest caveat: V4’s tool-call support still lags. At high-volume agentic workloads where reliability matters more than cost, it requires more defensive error handling than Claude or GPT-5.5. For budget routing layers and draft generation, it’s a strong choice. For mission-critical agents where a failed tool call causes user-facing errors, I’d stay on Claude Sonnet 4.6 or GPT-5.5.

The 90× cost gap that’s reshaping production AI

The pricing spread between frontier and budget-tier models is now the widest it’s ever been. This isn’t a reason to use the cheapest option — it’s a reason to build routing logic instead of treating your model choice as a fixed infrastructure decision.

Where models become products: Cursor, Codex, Claude Code

The model benchmarks matter, but for most developers the interface is where you spend time. Cursor, Claude Code, and OpenAI Codex have all shipped meaningful updates this cycle, and they’re increasingly diverging in philosophy as much as capability.

Beyond the chatbot: Perplexity Comet, Copilot Wave 3, and Figma AI

Perplexity Comet — Personal Computer Probable impact

Perplexity’s “Personal Computer” feature launched inside Comet for desktop and is expanding to Mac (and soon Windows). It extends their Computer-use capability to local file editing, local browsing, and voice orchestration — meaning you can dictate a research task, have Perplexity open tabs, read PDFs, and synthesize findings, all while you do something else. Claude Opus 4.7 is now the default orchestrator model for these multi-step workflows, with GPT-5.4 and Claude Sonnet 4.6 available as alternatives. Voice mode runs on OpenAI’s GPT-Realtime 1.5, with “over 25% more reliable interactions” compared to the previous version.

What I haven’t tested: how it handles documents with sensitive data when the local file system is involved. For anyone using this with confidential materials, that’s worth reading their data handling documentation carefully before deploying in a professional context.

Microsoft Copilot Wave 3 + Agent 365 Established

Microsoft launched Copilot Wave 3 on May 1, along with Agent 365 — a governance and security control plane for enterprise agents. The new E7 “Frontier Suite” license at $99/user/month bundles E5, Copilot, Agent 365, and Entra Suite. That’s Microsoft’s answer to the “agent sprawl” problem at large organizations: one license, one control plane, one audit trail.

The technical detail worth noting: Copilot Wave 3 explicitly orchestrates across Claude, GPT, and Microsoft’s own models depending on the task. SharePoint agent creation is in public preview, powered by Claude. This is the first Microsoft enterprise product to ship with Anthropic’s model as a named component — a meaningful signal about where the partnership is heading.

Figma AI — May Updates Established

Figma shipped meaningful AI updates across their entire product line in May. Voice-to-text in Figma Make lets you dictate prompts directly in the design agent chat. “Question cards” present structured choices before Make moves forward — reducing the back-and-forth that’s plagued AI design tools since they launched. FigJam got cell merging in tables, improved arrow routing, and a recenter button. Draw got a dedicated “text on a path” tool and brush/texture controls with directional noise. None of these are headline features, but collectively they suggest Figma is treating AI as infrastructure rather than a demo — which is the right instinct.

The “two years behind” argument is empirically wrong now

Six months ago, the standard take was that open-source models lagged the frontier by roughly 18–24 months. That argument has collapsed. GLM-5.1 briefly topped SWE-bench Pro in April — the first open-weight model ever to do so. Kimi K2.6 from Moonshot AI hit Tier A (87/100) on independent coding benchmarks. DeepSeek V4-Pro is within 7–8 percentage points of Claude Opus 4.7 on SWE-bench Verified.

Three models worth watching:

• DeepSeek V4-Flash — $0.14/M, MIT license, 1M context. The cheapest capable model ever released publicly. Use it for routing layers and first-pass drafts.
• Kimi K2.6 (Moonshot AI) — Full API access landed in early May. Competitive coding performance at lower price points than Western frontier models. Worth testing if you’re cost-sensitive on coding tasks.
• Qwen 3.5 — Alibaba’s latest, Tier C (55/100) on independent benchmarks with the right setup. For local inference on an RTX 5090 or M4 Ultra, it’s the closest open alternative to a production coding agent. Not a substitute for Claude. Hobbyist expectations, not production ones.

The practical caveat on running open-source locally: the toolchain failure rate is still real. Ollama lifecycle issues, missing tool-call parsers, cloud timeout problems — half the challenge of running open-source in 2026 is the stack that runs the model, not the model itself.

What I’d actually do this month

April and May 2026 demonstrated something that’s been building for two years: the frontier isn’t a place you arrive at and hold. It’s a moving target that now shifts week-to-week. DeepSeek V4 appeared in 24 hours. Kimi K2.6 went from preview to full API in days. Sonnet 4.8 is reportedly landing in May. Whatever model you’re using as your default today will be outdated by the time this post is a month old.

That’s not a counsel of despair — it’s a design prompt. Build your pipelines to swap models without rewriting your core logic. Keep a routing layer. Eval before you commit. And if you’re not running even a basic cost comparison between models for your actual workloads, that’s probably where to start this week.