← All digests
AI Developer Digest

Fri, Jul 24, 2026

9 signals that cleared the gate41 scanned25 min read
The Signal — start here
Claude Opus 5 is the story today — Anthropic launched their new flagship agentic model with benchmark numbers that are hard to ignore: 43.3% on Frontier-Bench v0.1 (2.3x Opus 4.8, ahead of Fable 5's 33.7%), 97.0% on SWE-bench Verified (prior SOTA was Mythos 5 at 95.5%), 1,861 Elo on GDPval-AA v2 (above Fable 5's 1,747 and GPT-5.6 Sol's 1,736), and — most significantly — 30.2% on ARC-AGI-3, more than tripling any prior model's score on a benchmark explicitly designed to resist memorization. All at $5/$25 per MTok, unchanged from Opus 4.8. There are real breaking changes that require code review: thinking is on by default, max_tokens now caps the sum of thinking tokens + response tokens, and thinking: {"type": "disabled"} with effort xhigh or max returns a 400 error. Separately, today is the hard cutoff for the DeepSeek V4 legacy API alias retirement — deepseek-chat and deepseek-reasoner stop accepting requests at 15:59 UTC.
Must-reads today
1
Claude Opus 5 — SOTA on ARC-AGI-3 (30.2%, >3x prior record), Frontier-Bench (43.3%), SWE-bench Verified (97.0%), and GDPval-AA at Opus prices; two breaking API changes require code review before migration
2
DeepSeek V4 alias retirement — deepseek-chat and deepseek-reasoner hard-retired today at 15:59 UTC; migrate to deepseek-v4-flash / deepseek-v4-pro now or integrations break
3
Claude Opus 4.7 fast mode removed — speed: "fast" on claude-opus-4-7 now returns a 400 error; migrate to Opus 4.8 or Opus 5 fast mode

Breaking Changes

3
●Breaking

Claude Opus 5: Disabling Thinking with `xhigh`/`max` Effort Now Returns 400

What changed
On Claude Opus 4.8, thinking: {"type": "disabled"} was accepted at any effort level. On Claude Opus 5, passing thinking: {"type": "disabled"} with effort xhigh or max returns a 400 error immediately.
TL;DR
Claude Opus 5 hard-rejects requests that disable thinking at effort xhigh or max — a breaking change from Opus 4.8 that fails immediately on migration.
Developer signal
Two paths to resolve before migrating to Opus 5: (1) Keep thinking disabled → lower the effort level. thinking: {"type": "disabled"} is still valid at low, medium, and high. Set output_config: {"effort": "high"} if you need the highest allowed effort with thinking off. (2) Keep the effort level → remove the thinking field. Opus 5 defaults to adaptive thinking, so removing the thinking field achieves the same result as {"type": "adaptive"} and works at any effort level. Separately, because thinking is now on by default, your existing max_tokens settings — sized for response-only output on Opus 4.8 — may now be too tight. Thinking tokens count against max_tokens. Run test requests and watch for stop_reason: "max_tokens" when you didn't expect truncation. Anthropic recommends 64,000+ max_tokens for xhigh/max effort sessions. Also watch for model behavior changes that require no code changes but affect output: Opus 5 writes longer default responses, narrates progress in agentic sessions more often, delegates to subagents more readily, and self-verifies — remove any existing "include a final verification step" or "use a subagent to verify" instructions; they cause over-verification on Opus 5.

Affects you ifYou call Claude Opus 5 with thinking: {"type": "disabled"} and effort xhigh or max; or you migrate from Opus 4.8 with max_tokens values tuned for non-thinking output.EffortModerate — code review required; the 400 error surfaces immediately on first test request.
●Breaking

Claude Opus 4.7 Fast Mode Removed

What changed
Fast mode for claude-opus-4-7 is removed as of today (June 25 deprecation, 30-day window). Requests to claude-opus-4-7 with speed: "fast" now return a 400 error.
TL;DR
claude-opus-4-7 + speed: "fast" returns a 400 error as of today; migrate to Opus 4.8 or Opus 5 fast mode.
Developer signal
Update model to claude-opus-4-8 (or claude-opus-5) and keep speed: "fast". Fast mode for Opus 4.8 and Opus 5 are both in research preview on the Claude API. Fast mode for Opus 5 is priced at $10/$50 per MTok and is API-only — not available on Bedrock, Google Cloud, or Foundry. If you are not on the fast mode waitlist, remove speed: "fast" from requests and run at standard speed; join the waitlist at claude.com/fast-mode.

Affects you ifYou call claude-opus-4-7 with speed: "fast" — these requests now return a 400 error.EffortQuick — model ID update and optional fast mode waitlist enrollment.
●Breaking

DeepSeek V4 Legacy API Aliases Retired — Hard Cutoff Today at 15:59 UTC

What changed
The deepseek-chat and deepseek-reasoner model aliases are permanently retired as of July 24, 2026, 15:59 UTC. During the grace period (since DeepSeek V4 GA on July 19-20), these aliases were transparently routed to deepseek-v4-flash. After the cutoff, requests using either alias return an error — no silent fallback.
TL;DR
deepseek-chat and deepseek-reasoner stop accepting requests at 15:59 UTC today; migrate model IDs to deepseek-v4-flash / deepseek-v4-pro or your integration breaks without warning.
Developer signal
Update model ID strings before 15:59 UTC: deepseek-chat → deepseek-v4-flash (284B total params / 13B active, 1M context, the faster/cheaper variant); deepseek-reasoner → deepseek-v4-pro (1.6T total / 49B active, 1M context, 384K max output, the higher-reasoning variant). The routing was already live in shadow mode: if you called deepseek-chat during the grace period, you were already getting V4-Flash responses. The change is only in model ID string validation — the responses have been V4 since July 19. Validate your response parsing after the swap: if your integration relied on any response format differences between the old generation and V4, catch those now. DeepSeek V4 pricing has peak/off-peak tiers; V4-Flash runs as low as $0.14/MTok input off-peak. If you use LiteLLM, the model aliases may need to be updated in your router config as well.

Affects you ifYou call the DeepSeek API with model string "deepseek-chat" or "deepseek-reasoner" — these fail after 15:59 UTC today.EffortQuick — string replacement in model configuration; validate response behavior after swap.
DeepSeek API Docs / multiple confirmations | Date: July 24, 2026 | Link: https://api-docs.deepseek.com/news/news260424/https://api-docs.deepseek.com/news/news260424/

Model Releases

1
High

Claude Opus 5: New SOTA on ARC-AGI-3, Frontier-Bench, SWE-bench Verified, and GDPval-AA at Opus Pricing

What changed
Claude Opus 5 (claude-opus-5) launches as Anthropic's primary agentic model. Key deltas from Opus 4.8: (1) 1M token context window is now the default and the maximum — no smaller context variant; (2) thinking is on by default at all effort levels; (3) full effort ladder now includes max level, no beta header; (4) prompt cache minimum lowered from 1,024 to 512 tokens; (5) mid-conversation tool changes in beta (mid-conversation-tool-changes-2026-07-01); (6) new SOTA across four benchmarks (see Benchmarks section); (7) available in GitHub Copilot Enterprise/Business. Priced identically to Opus 4.8 at $5/$25 per MTok; fast mode research preview at $10/$50 per MTok (API only).
TL;DR
Claude Opus 5 sets new SOTA on ARC-AGI-3 (30.2%, >3x prior record), Frontier-Bench v0.1 (43.3%, beats Fable 5), SWE-bench Verified (97.0%), and GDPval-AA v2 (1,861 Elo) at $5/$25 per MTok — same as Opus 4.8.
Developer signal
Concrete action items for migrating: (1) Model ID: replace claude-opus-4-8 with claude-opus-5 in SDK/API config. (2) Thinking on by default: no thinking field needed; remove explicit thinking: {"type": "adaptive"} calls, they're redundant. Cannot disable thinking at xhigh/max effort (returns 400 — see Breaking Changes). (3) Raise max_tokens: thinking tokens count against the cap; set 64k+ for xhigh/max effort tasks. (4) Remove verification instructions: delete "include a final verification step" or "use a subagent to verify" from prompts — Opus 5 self-verifies and will over-verify if asked. (5) New max effort level: available with no beta header; enables the deepest reasoning at the cost of more tokens; use for capability-critical tasks. (6) Cache: minimum cacheable prefix is now 512 tokens (down from 1,024) — short system prompts that couldn't cache on Opus 4.8 may now cache automatically. (7) Mid-conversation tool changes (beta): add anthropic-beta: mid-conversation-tool-changes-2026-07-01 to enable adding/removing tools between turns while preserving the prompt cache — enables dynamic tool graphs in agentic pipelines (see SDK section for tool_change events). (8) GitHub Copilot: Opus 5 is live in Copilot Enterprise/Business; admins must enable the Claude Opus 5 policy in Copilot settings before users can select it. (9) Fallbacks default mode: add anthropic-beta: server-side-fallback-2026-07-01 and pass fallbacks: {"type": "default"} to use Anthropic's recommended refusal-category-based routing without maintaining your own fallback list.

Affects you ifYou build on Claude Opus 4.8 and are evaluating migration; you need SOTA agentic coding performance; you use GitHub Copilot Enterprise or Business; you have existing prompts with tight max_tokens settings or explicit verification instructions.EffortModerate — model ID change is Quick, but thinking-on-by-default and max_tokens behavior change require testing existing prompts before production rollout.

API & SDK Changes

1
Medium

Anthropic SDK Python v0.120.0: Claude Opus 5, Tool Change Events, Server-Side Fallbacks Default

What changed
Three additions over v0.119.0: (1) claude-opus-5 added as a typed model constant; (2) tool addition/removal blocks and tool_change events added — the event type that surfaces mid-conversation tool management per turn; (3) client-side fallback credit token types expanded and a new "default" mode added for server-side fallbacks, which applies Anthropic's recommended routing by refusal category.
TL;DR
Anthropic SDK Python v0.120.0 adds claude-opus-5 support, tool_change event handling for mid-conversation tool addition/removal, and a "default" mode for server-side refusal fallbacks.
Developer signal
pip install --upgrade anthropic to pick up v0.120.0. Three use cases: (1) Opus 5 API calls: model="claude-opus-5" now resolves to the correct typed constant; without upgrading, the string "claude-opus-5" still works but you lose type safety in SDK method signatures. (2) Tool change events: when using the mid-conversation-tool-changes-2026-07-01 beta header, the event stream now includes tool_change events surfacing which tools were added or removed per turn — critical for building orchestrators that track tool lifecycle without polling. (3) Fallbacks default mode: pass fallbacks={"type": "default"} with the server-side-fallback-2026-07-01 beta header to let Anthropic's recommended fallback model routing (by refusal category) apply automatically. Useful as a starting-point policy before you tune your own fallback list.

Affects you ifYou use the Anthropic Python SDK and are migrating to Claude Opus 5; you build multi-turn agents that add/remove tools between turns; you use or plan to use server-side refusal fallbacks.EffortQuick — pip install --upgrade anthropic.

Research

Nothing cleared the quality gate this period. arXiv cs.AI and cs.CL search returned active papers but none from recognized labs with code repositories and concrete benchmark numbers from July 23-24. HuggingFace Papers daily returned 403 on direct fetch. Google DeepMind Publications showed last notable entry July 2 (LLM overthinking analysis). No qualifying papers from this window.

Tooling

1
Notable

llama.cpp b10105: `--mlock`, `--mmap`, `--directio` Consolidated into `--load-mode`

What changed
PR #20834 consolidates three memory management CLI arguments (--mlock, --mmap, --directio) into a unified --load-mode parameter. The three legacy flags remain functional with deprecation warnings. The + sign now functions as an actual argument modifier; inverse commands now perform their actual opposite operation rather than doing nothing; mlock activation bugs corrected.
TL;DR
llama.cpp b10105 unifies --mlock, --mmap, and --directio into --load-mode, deprecating the individual flags with backward-compatible warnings and fixing associated mlock activation bugs.
Developer signal
If you have shell scripts or systemd/startup commands using --mlock, --mmap, or --directio for memory management, they continue working but emit deprecation warnings. Migrate to --load-mode with the appropriate value to silence warnings. No behavior change in this release for correctly functioning setups — but the mlock bug fix means previously broken --mlock invocations may now correctly lock memory, which could affect systems where --mlock was specified but silently not applying. Verify that your memory-locked inference setups still behave as expected after upgrading to b10105.

Affects you ifYou use llama.cpp CLI with --mlock, --mmap, or --directio in inference scripts; or you expected --mlock to be active on prior builds (the bug fix may change behavior).EffortQuick — flag migration in scripts; backward compat maintained with deprecation warnings.

Benchmarks & Leaderboards

1
High

Claude Opus 5 Sets SOTA on ARC-AGI-3 (30.2%), Frontier-Bench (43.3%), SWE-bench Verified (97.0%), and GDPval-AA v2 (1,861 Elo)

What changed
Four leaderboard positions changed on Claude Opus 5 launch:

  • ARC-AGI-3: Opus 5 at 30.2% — prior best was ~7-8% (GPT-5.6 Sol); this is more than tripling the previous record

  • Frontier-Bench v0.1 (agentic terminal coding): Opus 5 at 43.3% — new #1 ahead of Fable 5 (33.7%) and well above Opus 4.8 (18.7%)

  • SWE-bench Verified: Opus 5 at 97.0% — new #1 ahead of Mythos 5 (95.5%), Fable 5 (95.0%), and Opus 4.8 (88.6%)

  • GDPval-AA v2 (knowledge-work quality vs expert humans): Opus 5 at 1,861 Elo — above Fable 5 (1,747) and GPT-5.6 Sol (1,736)

TL;DR
Claude Opus 5 takes #1 across ARC-AGI-3, Frontier-Bench, SWE-bench Verified, and GDPval-AA on launch day — notably ahead of Fable 5 (Anthropic's frontier model) on three of the four benchmarks, at half the price.
Developer signal
Four benchmarks, four signal levels: (1) ARC-AGI-3 is the most meaningful signal here — François Chollet designed it to resist memorization and measure novel reasoning. Tripling the prior best is not a marginal improvement; it represents a qualitative jump in novel problem-solving. This is the test-time compute scaling story made concrete. (2) Frontier-Bench v0.1 (agentic terminal coding) is the most directly applicable signal for developers building agentic coding pipelines: Opus 5 (43.3%) beats Fable 5 (33.7%) on the benchmark that most directly measures autonomous code task completion. For teams choosing between Fable 5 and Opus 5 for coding agents, Opus 5 is the stronger choice at half the price per token. (3) SWE-bench Verified at 97.0% has a known contamination caveat: OpenAI's internal audit found some SWE-bench Verified tasks overlap with training data, so the number is not purely a measure of novel bug-fixing ability. Treat it as a directional signal rather than a hard capability ceiling. (4) GDPval-AA v2 at 1,861 Elo tests structured knowledge-work quality against expert human baselines; the gap over Fable 5 (114 Elo points) is wider than is typical between adjacent model generations, suggesting a more significant jump in professional task quality.

Affects you ifYou choose models based on coding, agentic, or knowledge-work benchmarks; you maintain internal model performance baselines.EffortN/A — leaderboard data.
Multiple benchmark trackers (officechai.com, morphllm.com, vals.ai) | Date: July 24, 2026 | Link: https://officechai.com/ai/claude-opus-5-arc-agi-3/ (ARC-AGI-3); https://www.morphllm.com/claude-benchmarks (Claude benchmarks overview)https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5 (Anthropic's capability summary); https://officechai.com/ai/claude-opus-5-arc-agi-3/ (ARC-AGI-3)

Technical Discussions

Nothing cleared the quality bar this period.

Quick Hits

Worth Watching (Announced, Not Yet Shipped)

ItemETANotes
Kimi K3 open weightsJuly 27, 2026 (3 days)Moonshot confirmed weights under Modified-MIT-style license. MXFP4 quantization for Blackwell planned. 2.8T total / ~32B activated. Covered in July 18 digest.
MCP 2026-07-28 final specificationJuly 28, 2026 (4 days)Removes initialize/initialized handshake, Mcp-Session-Id. Adds Mcp-Method and Mcp-Name headers. Error code for missing resources changes from -32002 to -32602. Tier 1 SDKs in beta. Covered in July 18 digest.
Anthropic Workbench sunset + experimental prompt tools retirementAugust 17, 2026/v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt will error after August 17. Export saved prompts before then.
Claude Opus 4.1 retirementAugust 5, 2026Migrate to Opus 4.8 or Opus 5.
OpenAI o3 retirementAugust 26, 2026Migrate to GPT-5.6 Sol/Terra/Luna.
Imagen 4 shutdownAugust 17, 2026Migrate to Imagen 4 Ultra or Veo 3.

Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.