Skip to main content
Fleece AI — the gold ram mascotFleece AI
Book a demoGet started free
Back to blog
GuideExplainers & Guides8 min readSeptember 28, 2026

Anthropic Updates April 2026: Claude Opus 4.7, GPT-5.3-Codex & AI Releases

By Loïc Jané · Founder, Fleece AI

April 2026 AI Releases: What Changed for Workflows

At a Glance: April 2026 shipped four releases that matter for workflow automation — Claude Opus 4.7 (Anthropic's new flagship, April 16), GPT-5.3-Codex and Codex-Spark (OpenAI's agentic coding models), ChatGPT Agent Mode, and Claude Sonnet 4.6 with 1M-token extended thinking. This roundup explains each release, benchmarks the real impact, and maps it onto Fleece AI's Business tier. Published April 17, 2026; updated September 28, 2026.

Update (September 2026): this is a snapshot of April 2026. Several of these models have since been superseded: Anthropic followed Opus 4.7 with Claude Opus 4.8, Opus 5 and, on September 22, 2026, Claude Opus 5.5; Claude Sonnet 5 replaced Sonnet 4.6 as the balanced model; and OpenAI released GPT-6 Astra (September 3) and GPT-6 Sol and Luna (September 22). On Fleece AI, the default model on every plan is now GLM 5.3 Flash, and the Business plan offers Claude Opus 5.5, 4.8 and 4.7; Opus 4.6 is no longer selectable for new agents (agents already pinned to it keep running). The Fleece AI recommendations below describe April 2026.

The April 2026 AI releases landed in a two-week window and collectively tighten the gap between model benchmarks and production-ready agents. If you automate work with AI, the delta between March and April 2026 is larger than the entire Q1 in raw capability. The frontier models race continues to accelerate, with Anthropic and OpenAI both shipping meaningful upgrades in the same window. This guide breaks down each model release and what it concretely changes for your workflows.

ReleaseReleasedHeadline changeWhat it means for Fleece AI
Claude Opus 4.7April 16, 2026+13% coding benchmarks, 3.75-MP vision, and self-checking at the same $5/$25 per million tokensRecommended Anthropic model on Business in April 2026; still offered, next to Opus 5.5
GPT-5.3-Codex and Codex-SparkLate March to mid April 2026OpenAI's most capable agentic coding models at the time, combining Codex and GPT-5 training stacksNot offered on Fleece AI
ChatGPT Agent ModeBroad rollout in Q1 2026Visual browser, terminal, and connectors, capped at 40-400 messages/month with no scheduled executionComplementary for interactive one-offs, not scheduled multi-agent workflows
Claude Sonnet 4.6Q1 2026Balanced model with extended thinking and a 1M-token context window (beta), fewer tokens per taskNot offered on Fleece AI

Table of Contents

Key Takeaways

  • Anthropic released Claude Opus 4.7 on April 16, 2026 — a strict upgrade over Opus 4.6 with +13% coding benchmarks and 3.75-MP vision at the same price point, per Anthropic's announcement (April 2026).
  • OpenAI shipped GPT-5.3-Codex and Codex-Spark — the most capable agentic coding models to date, combining Codex and GPT-5 training stacks per OpenAI's announcement (April 2026).
  • ChatGPT Agent Mode went broadly available on Pro, Plus, and Team, with a visual browser and connectors — but is capped at 40 messages/month on Plus and 400 on Pro per OpenAI's Help Center (2026).
  • Claude Sonnet 4.6 became Anthropic's balanced option — extended thinking, 1M-token context (beta), fewer tokens consumed per task.
  • For Fleece AI users, the practical April 2026 upgrade is one line: switch Business agents from Opus 4.6 to Opus 4.7 and expect more tasks resolved on the first try.

1. Claude Opus 4.7

Released: April 16, 2026. Pricing: $5 / $25 per million tokens (unchanged from Opus 4.6). Headline improvements: +13% coding benchmarks, 3.75-MP vision input, self-checking behavior, updated tokenizer, longer output tokens per task, and safety safeguards against prohibited cybersecurity use.

According to Anthropic's release notes (April 2026), Opus 4.7 beats Opus 4.6, GPT-5.4, and Gemini 3.1 Pro on a number of key benchmarks. Axios reported (April 2026) that Anthropic openly conceded Opus 4.7 is "less broadly capable" than their gated Mythos Preview — meaning a stronger model exists, but is not generally available.

Impact on Fleece AI workflows (April 2026): Opus 4.7 became the recommended Anthropic model on Business plans, and agents previously pinned to Opus 4.6 kept running. As of September 2026, Business also offers Opus 4.8 and Opus 5.5, and the default model for new agents on every plan is Kimi K3. The self-checking behavior is the single most useful change for unattended autonomous agents — fewer confidently-wrong answers at 3 AM.

2. GPT-5.3-Codex and Codex-Spark

Released: Late March to mid April 2026. Headline: Most capable agentic coding models OpenAI has shipped, combining Codex and GPT-5 training stacks.

According to OpenAI's announcement (April 2026), GPT-5.3-Codex-Spark is designed for end-to-end coding agents — code generation, reasoning, tool use, and long-running execution in a single unified model. Codex-Spark specifically targets the "agent builds a feature, tests it, commits, and opens a PR" workflow pattern.

OpenAI also introduced GPT-5.3 Instant Mini — a fallback model that kicks in when users hit their rate limit on GPT-5.3 Instant. Per OpenAI's release notes, Instant Mini feels "more natural in conversation, with stronger writing and contextual awareness."

Impact on Fleece AI workflows: OpenAI is clearly pushing the "agent coder" frontier. For teams using Fleece AI to automate GitHub workflows, the gap between a scripted automation and an agent that writes its own code is shrinking fast. As of April 2026, Fleece AI's default model was GPT-5.2. Codex-Spark has not been added; the default is now GLM 5.3 Flash, and GPT-5.2 remains available on every plan.

3. ChatGPT Agent Mode

Released: Broad rollout to Pro, Plus, and Team in Q1 2026; Enterprise and Education followed. Headline: Agentic capabilities embedded in ChatGPT — visual browser, text browser, terminal, API access, and connectors (Gmail, GitHub, Google Drive).

Per OpenAI's Help Center (2026), ChatGPT Agent is activated from the tools dropdown in the composer. Usage is capped at 400 messages per month for Pro, 40 for other paid tiers, with credit packs available for overages.

ChatGPT Agent Mode excels at interactive, one-off tasks — scraping a comparison, filling a form, drafting a reply with Gmail context. It does not support scheduled execution: the session must be open and the user must trigger each run.

Impact on Fleece AI workflows: ChatGPT Agent Mode competes with Fleece AI only on the "interactive one-off" job. For the scheduled, multi-agent, multi-app workflows Fleece AI is built for, the two are complementary rather than overlapping. For the full side-by-side see our ChatGPT Agent Mode vs Fleece AI comparison.

4. Claude Sonnet 4.6

Released: Q1 2026. Headline: Anthropic's balanced model — stronger agentic search, fewer tokens consumed, extended thinking, 1M-token context window (beta).

Sonnet 4.6 is positioned between Haiku (fast/cheap) and Opus (deep reasoning). Like claude code, it is designed for practical developer workflows where cost per task matters. It supports extended thinking — the model can allocate extra compute to hard problems without switching to a different endpoint. The 1M-token context window (in beta) makes Sonnet competitive for long-document workflows that previously required Opus.

Impact on Fleece AI workflows: Fleece AI does not offer Sonnet models, but the underlying shift matters: Anthropic is widening the gap between "balanced" and "flagship," giving teams more surgical control over cost vs capability. Sonnet 4.6 has since been followed by Claude Sonnet 5 ($2/$10 per million tokens).

5. Honorable Mentions

Claude Mythos Preview

Anthropic's gated research model, accessible only through Project Glasswing for defensive cybersecurity work. Mythos outperforms Opus 4.7 but was invitation-only in April 2026. As of September 2026, Claude Mythos 5 and 5.1 are still offered by invitation only, through Project Glasswing.

GPT-5.4-Cyber

A fine-tuned variant of GPT-5.4 with fewer restrictions on sensitive cybersecurity tasks (vulnerability research, analysis). Tiered access via OpenAI — niche for most Fleece AI users, relevant for security teams.

GPT-5.3 Instant Mini

The cheaper fallback when users hit GPT-5.3 Instant rate limits. Not a flagship release but quietly improves the ChatGPT experience for heavy users. Meanwhile, the open source model space saw significant progress, though most releases did not impact enterprise agent workflows.

What Shipped Since: Three Months of Adoption

The four releases above have now been live long enough to judge. What our own post-release analyses show:

  • Opus 4.7 became the working Anthropic model over the summer. Business agents moved to 4.7, while agents pinned to 4.6 kept running unchanged. It has since been followed by Opus 4.8 and Opus 5.5 on the same plan. The self-checking behavior we flagged in April proved to be the most consequential change for unattended runs: fewer confidently-wrong answers on scheduled overnight work.
  • Codex-Spark stayed off Fleece AI. The loop it targets — an agent that builds a feature, tests it, commits, and opens a PR — is exactly the workflow we want, but it was not added to the model list.
  • ChatGPT Agent Mode settled into its niche. Its message caps and single-session design confirmed the April prediction: it complements interactive one-offs, not scheduled multi-agent workflows.
  • Sonnet 4.6 found its tier, then gave way to Sonnet 5. Extended thinking with a 1M-token context window made it Anthropic's balanced option until Claude Sonnet 5 replaced it.

If you adopted these releases in April, the three-month mark is the right time to re-benchmark your own workflows against current numbers before committing to a switch.

What This Means for Fleece AI Workflows

The April 2026 AI releases compress into three practical actions for Fleece AI users:

  1. Switch Anthropic agents from Opus 4.6 to Opus 4.7 on Business. Same price, strictly better, fewer failures on long-running agents. (September 2026: Claude Opus 5.5 is now also on Business, at a lower $4/$20 per million tokens on Anthropic's API.)
  2. Reconsider "ChatGPT is good enough" mental models. ChatGPT Agent Mode is real but session-bound and rate-capped. For scheduled business automation, Fleece AI's delegative model remains the right tool.
  3. Watch the agentic coding space. GPT-5.3-Codex-Spark and Opus 4.7 together move agents from "generate code" to "ship features" territory. If you automate engineering work, re-evaluate your stack quarterly.

For a broader model comparison see best AI models for workflow automation 2026.

Ready to run Claude on real workflows? Upgrade to Fleece AI Business — 3,000+ integrations, scheduled agents, and Claude Opus 5.5, 4.8 and 4.7.

Frequently Asked Questions

Is Claude Opus 4.7 worth switching to from Opus 4.6?

Yes for most workflows. Opus 4.7 is +13% better on coding benchmarks, has 3.75-MP vision (vs ~1 MP on 4.6), and introduces self-checking — all at the same $5/$25 per million tokens. Agents pinned to 4.6 keep running. As of September 2026, Fleece AI Business also offers Claude Opus 5.5, which is priced lower ($4/$20) on Anthropic's API. See the Opus 4.7 review for benchmarks.

How does ChatGPT Agent Mode compare to Fleece AI for business automation?

ChatGPT Agent Mode runs inside a single chat session and is capped at 40-400 messages per month. Fleece AI runs autonomous agents on cron schedules across 3,000+ apps. For one-off research, ChatGPT Agent is faster. For recurring automation, Fleece AI is the right platform. Full comparison in our ChatGPT Agent Mode vs Fleece AI article.

What is Claude Mythos and when will it be available?

Mythos is Anthropic's gated model family, accessible only through Project Glasswing for defensive cybersecurity work. As of September 2026, Claude Mythos 5 and 5.1 are listed on Anthropic's pricing page as limited availability, by invitation. Anthropic has not announced general availability. The strongest generally available Anthropic models are now Claude Fable 5.1 and Claude Opus 5.5.

Does Fleece AI support GPT-5.3-Codex-Spark?

No. As of September 2026, Fleece AI offers GLM 5.3 Flash (the default), GPT-5.2, Mistral Medium 3.5, GPT-5.6 Luna, Qwen 3.8 Max, Kimi K3 and DeepSeek V4.1 Flash on Starter; adds GPT-5.6 Terra on Pro; and adds GPT-5.6 Sol and Claude Opus 5.5, 4.8 and 4.7 on Business. We evaluate new models based on real-world agent reliability, not just benchmarks.

What should I change in my Fleece AI workflows after April 2026?

Three things. First, set Business agents to Opus 4.7 instead of 4.6. Second, re-check any agent that has been pinned to an old model version since Q1. Third, revisit your multi-agent hierarchy: the reliability bump on 4.7 may let you simplify chains that previously needed manual checkpoints.

The Bottom Line

April 2026 was an unusually dense month for AI releases. Anthropic's cadence held: Claude Opus 4.8, Opus 5 and, on September 22, Opus 5.5 followed. The single most valuable action for Fleece AI users is mechanical: switch Business-tier workflows to Claude Opus 4.7. Everything else — ChatGPT Agent Mode, Codex-Spark, Sonnet 4.6 — either complements Fleece AI for niche jobs or validates the delegative-agent thesis from a different angle.

Related Articles

Start your trial on Fleece AI — Kimi K3, GPT-5.6, Claude Opus 5.5 (Business), 3,000+ integrations, and scheduled agents that run while you sleep.

Ready to delegate your first task?

Deploy your first AI agent in under 60 seconds.

Get started free

Related articles

Agent IA pour PME : automatiser sans recruter (2026)

7 min read

AI Agents for SMEs: Scale Without Headcount (2026)

7 min read

Agent IA pour TPE et indépendants (2026)

6 min read

Anthropic Updates April 2026: Claude Opus 4.7, GPT-5.3-Codex & AI Releases | Fleece AI