Round 1 closed on a line worth repeating: capability and access are being negotiated separately, and every architecture needs to survive either one changing without warning. This round is the proof. In the space of five days, three frontier labs shipped their most consequential releases of the summer — and the resulting pressure forced Anthropic into a public, twice-repeated climbdown on a policy it had announced with confidence just two weeks earlier.
None of this happened in a vacuum. SpaceXAI's Grok 4.5 landed July 8 trained directly on Cursor's developer session data. OpenAI answered July 9 with GPT-5.6 going fully public and merging ChatGPT with Codex into one desktop app. Anthropic, watching both, extended Fable 5's included access on paid plans twice in six days — first to July 12, then again to July 19 — while quietly resetting Claude Code's rate limits. Here's the full read on the week of July 6–13, the practical technique for running two frontier models in one loop, and what to actually build with while the ground keeps moving.
01 · FRONTIER
GPT-5.6 goes fully public — and OpenAI merges ChatGPT with Codex.
JULY 9GENERAL AVAILABILITYCHATGPT WORK LAUNCHES
Three weeks after previewing GPT-5.6 to roughly 20 trusted partners under government coordination, OpenAI shipped it globally on July 9 — across ChatGPT, Codex, and the API. The rollout is bigger than a model update: OpenAI merged its Codex desktop app and ChatGPT desktop app into a single application, and introduced ChatGPT Work, a new agent built specifically for long, multi-step office work.
What ChatGPT Work actually does
OpenAI's own framing: it can "take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work." One example from the launch: turn customer research into a campaign brief, generate marketing assets from it, and adapt them for different markets — all in one continuous run, with context preserved across every step. It connects to Google Drive, Slack, and Salesforce out of the box, uses Scheduled Tasks to keep projects moving even when you're away, and can turn messages from Teams or Slack directly into updated docs or slides.
PRODUCT STRUCTURE
One App: Chat, Work, Codex
The new desktop app for Mac and Windows unifies three surfaces — everyday chat, Work for research and finished deliverables, and Codex for software development. Existing Codex app users update in place and keep their projects.
ROLLOUT
Pro/Enterprise/Edu First
Web and mobile access to ChatGPT Work starts with Pro, Enterprise, and Edu plans; Plus and Business follow within days. On desktop, it's live for every plan today, including Free.
MODEL ACCESS
Sol / Terra / Luna by Tier
Free and Go users get GPT-5.6 Terra. Plus, Pro, Business, and Enterprise can choose among Sol, Terra, and Luna with adjustable effort levels. Ultra mode (subagent orchestration) is available to Pro and Enterprise in Work, and Plus-and-up in Codex.
SITES
Interactive Web Apps, Public Beta
Describe what you want built, add content and constraints, review a private preview, then publish. Rolling out to Pro, Pro Lite, and Edu now, Plus following — not yet available in the EEA, Switzerland, or UK.
GPT-5.6 Sol is priced at $5/M input and $30/M output — identical to GPT-5.5, making it effectively a free capability upgrade for existing Sol-tier users. Terra runs $2.50/$15, Luna $1/$6. OpenAI highlights a genuine jump in design judgment: with only high-level direction, the model can inspect its own rendered output — not just the underlying code — and catch visual or functional issues before handing work back.
Reading between the announcements
Several independent observers noted that much of what ChatGPT Work does — scheduled tasks, plugins, computer use, cross-app connections — already existed across ChatGPT and Codex separately. The genuine news is the merger itself: OpenAI is consolidating around agentic UX as the default surface rather than chat, using Codex's browser and computer-use infrastructure as the substrate for a much wider audience. It's the same structural move Anthropic made months earlier going from Claude Code to Cowork — both labs have concluded that "chat, but longer" isn't the product; a persistent, file-and-app-aware agent is.
02 · FRONTIER
The model-war week: Grok 4.5 lands, and Anthropic extends Fable 5 twice.
JULY 8–133 LABS · 5 DAYSFABLE 5 EXTENDED TWICE
Context matters here more than any single release. Four days separated the three biggest model moves of the summer — and the sequence reads like direct cause and effect.
JULY 8
Grok 4.5 launches publicly. SpaceXAI's first model since acquiring Cursor, trained on real Cursor developer session data. $2/M input, $6/M output — Musk calls it "roughly comparable to Opus 4.7, but much faster." Live in Grok Build, Cursor (all plans), and the SpaceXAI console.
JULY 9
GPT-5.6 goes fully public + ChatGPT Work launches. OpenAI merges ChatGPT and Codex into one app. Sol/Terra/Luna reach GA across ChatGPT, Codex, and API.
JULY 7 → 12
First Fable 5 extension. Anthropic pushes the included-access cutoff from July 7 to July 12, hours before the original deadline, via an X post and an updated support article — no dedicated newsroom announcement.
JULY 12–13
Second extension, hours after the first expired. Anthropic extends again to July 19, keeps Claude Code's weekly rate limits 50% higher through the same date, and quietly resets rate limits across the board.
Grok 4.5, in the numbers that matter
Grok 4.5 doesn't lead everywhere, but at $2/$6 per million tokens — over 60% cheaper than Opus 4.8 or GPT-5.5 — it's landing within a few points of the frontier on tasks that matter most for agentic coding, while xAI reports roughly 4× better token efficiency than Opus 4.8 on the same evaluation. One caveat worth flagging: Cursor disclosed that an earlier Cursor codebase snapshot was accidentally included in Grok 4.5's training data, which may have given it an edge on Cursor-specific benchmarks — treat in-IDE comparisons with that in mind. Also not yet available in the EU at launch (targeted mid-July).
| Benchmark | Fable 5 (max) | Grok 4.5 | Opus 4.8 (max) |
|---|
| Terminal-Bench 2.1 | 84.3% | 83.3% | 78.9% |
| SWE-Bench Pro | 80.4% | 64.7% | 69.2% |
| SWE Marathon (pass@1) | 24.0% | <strong>29.0%</strong> | 26.0% |
Reading the Fable 5 pattern honestly
Anthropic has now moved this deadline three times in eighteen days: the original July 1 return, a same-day-of-deadline push from July 7 to July 12, and a second push — announced hours after the July 12 cutoff had already passed — to July 19. Anthropic hasn't publicly tied any of this to the Grok 4.5 or GPT-5.6 launches, but the timing is difficult to read any other way, and outside commentary (including Anthropic's own Claude Code team, per reporting) has confirmed the underlying driver is a genuine compute-capacity constraint, with a stated aim to make Fable 5 a standard, permanent plan inclusion once capacity allows — not yet a commitment. If you're building anything durable on Fable 5's inclusion terms specifically, budget for the credit-metered fallback ($10/M input, $50/M output) rather than assuming next week's terms hold.
03 · PRODUCTIVITY
ChatGPT for PowerPoint goes GA — worldwide, every plan.
JULY 6ALL PLANS · GLOBALFREE THROUGH AUG 6
After a May beta, ChatGPT for PowerPoint is now generally available across every ChatGPT tier — Free, Go, Plus, Pro, Business, Enterprise, Edu, Teachers, and K-12 — installed as a Microsoft 365 add-in from Home → Add-ins. It builds decks from notes, documents, spreadsheets, images, or an existing presentation, rewrites and restructures slides, and can answer questions about a deck's narrative, gaps, and likely audience objections — all while keeping output editable natively in PowerPoint.
What's actually new versus the beta
OpenAI is explicit about the limits: complex chart editing, animations, custom fonts, and advanced template handling aren't fully supported yet, and the tool "can make mistakes... and may change or delete content if a request is unclear" — the standing advice is to keep a copy of any deck before letting it make sweeping edits.
- Global GA, not regional beta — every plan tier, worldwide, no waitlist.
- Token-based credit pricing — same model as ChatGPT for Excel/Sheets, roughly 20–110 credits per task depending on size and complexity; free for Business and Enterprise through August 6, 2026, then metered from the workspace credit pool.
- Connected-app support — where enabled, pulls content directly from Outlook and SharePoint into slides.
- Skills and plugins — repeatable presentation workflows can be packaged and invoked with
@, the same plugin system now shared across Work and Codex.
For Conneqt pitch work
This is a genuine time-saver for first-draft assembly — turning a discovery call's notes into a first-pass client deck, or restructuring an existing Conneqt brand deck for a new prospect's specific pain points. Treat it as a drafting accelerant, not a final-polish tool: run the brand-compliance and Arabic-register pass by hand afterward, the same way you would with any first draft.
04 · TECHNIQUE
You can run Fable 5 and GPT-5.6 Sol in the same loop, right now.
WORKS TODAYCLAUDE CODE + CODEX PLUGIN~50% TOKEN COST
A pattern spreading fast among power users this week: instead of picking one frontier model per task, split the job between two — one that plans and judges, one that types. The setup is short enough to try this afternoon.
The setup
The logic: Fable 5 is the most expensive model either lab sells, and the least efficient use of that cost is line-by-line typing a model half its price can do just as well. By having Fable spend its budget only on the parts that genuinely need judgment — architectural decisions, plan decomposition, reviewing Codex's output against the original intent — and delegating mechanical implementation to GPT-5.6 Sol, early adopters are reporting roughly half the token bill for comparable output versus running Fable 5 solo end to end.
The broader lesson generalizes past this specific pairing: with Sonnet 5, GLM-5.2, Grok 4.5, and GPT-5.6's own tiers all now viable "typing" layers at a fraction of frontier cost, the emerging default architecture for hard agentic work is a cheap-fast layer for execution and an expensive-smart layer for judgment — not one model doing everything.
- Open Claude Code and install the Codex plugin.
- Switch your active model to Fable 5.
- Add one routing block to your
CLAUDE.md instructing Fable to orchestrate, plan, decompose, and review — and to delegate actual implementation to Codex running GPT-5.6.
The catch
Fable 5 costs 2× Opus 4.8 per token ($10/$50 vs $5/$25). This orchestration pattern only pays off on genuinely long, hard, multi-step tasks — codebase migrations, multi-file refactors, complex feature builds — where the judgment-vs-typing split is real. For quick edits or single-file changes, the coordination overhead of two models isn't worth it; just use one model directly. And remember Fable 5's inclusion terms are currently a moving target (see Section 02) — this pattern gets more expensive fast once you're past the 50% weekly allowance.
05 · LEARNING
12 free courses to actually master LLMs — not just prompt them.
Worth a standalone section this round: a curated list of free, self-paced courses covering the full stack from fundamentals to production agents circulated widely this week. Useful for onboarding new Conneqt hires or leveling up client-facing teams beyond prompt-level familiarity.
01
Cohere LLM University foundations
Semantic search, generation, classification, and embeddings — a full NLP-to-LLM curriculum mixing theory with hands-on exercises.
02
Hugging Face LLM Course foundations
Transformers, Datasets, Tokenizers, and Accelerate libraries — starts with tokenization basics, builds to fine-tuning and sharing models on the Hub.
03
Hugging Face AI Agents Course agents
Agent fundamentals through the Hugging Face ecosystem — tool use, planning, and evaluation for agentic systems.
04
Google / Kaggle 5-Day Gen AI Intensive foundations
A structured five-day sprint through generative AI fundamentals, run jointly by Google and Kaggle.
05
DeepLearning.AI Short Courses breadth
Andrew Ng's library of focused, hour-scale courses spanning prompt engineering, fine-tuning, and applied LLM techniques.
06
Hugging Face Context Course advanced
Context engineering specifically — how long-context windows, memory, and retrieval interact in production LLM systems.
07
Google / Kaggle 5-Day AI Agents Intensive agents
The agents-focused companion to the Gen AI Intensive — planning, tool use, and multi-step orchestration in five days.
08
DeepLearning.AI Retrieval Augmented Generation Course rag
RAG fundamentals — chunking, embeddings, retrieval quality, and grounding generation in external knowledge.
09
DeepLearning.AI Building Agentic RAG with LlamaIndex rag
Combines agentic reasoning with retrieval — building agents that decide when and what to retrieve rather than retrieving blindly.
10
Weights & Biases AI Academy production
Experiment tracking, evaluation, and observability for LLM systems moving from notebook to production.
11
LangChain Academy: Intro to LangGraph and Deep Agents agents
Graph-based agent orchestration — state machines, cycles, and multi-agent coordination patterns in LangGraph.
12
DeepLearning.AI AI Agents in LangGraph agents
A hands-on companion to the LangChain Academy track, building working agents step by step inside LangGraph.
For the Conneqt team
If onboarding a new hire onto the Conneqt Brain or Hermes stack, the fastest path is Hugging Face's LLM Course (foundations) → AI Agents Course (agent basics) → LangChain Academy's LangGraph track (orchestration patterns that map directly onto how Hermes and MoA presets work). All twelve are free and self-paced — no budget approval needed to start this week.
06 · TOOLS
The coding-tool landscape as of this week, plus a library of ready-made agent loops.
With three frontier models and a product merger landing in five days, it's worth a snapshot of where the coding-tool market actually stands — the models are moving fast, but the tools you point them through are where the real workflow decisions live.
Best AI coding tools, 2026
Open-source coding models worth self-hosting
FRONTIER-BACKED IDEs
Claude Code · OpenAI Codex
The two lab-native options — deepest integration with their own model families, first access to new capabilities (Fable 5 orchestration, GPT-5.6 ultra mode).
MODEL-AGNOSTIC EDITORS
Cursor · Windsurf · Cline · Aider
Route across Claude, GPT, Grok, and open models from one interface. Cursor now ships Grok 4.5 alongside Composer 2.5 post-SpaceX acquisition.
CLOUD-NATIVE AGENTS
Replit Agent · Amazon Q Developer
Full environment-to-deploy pipelines rather than editor plugins — the right fit when the deliverable is a running app, not just a diff.
ENTERPRISE + PLATFORM
GitHub Copilot · Sourcegraph Cody · Gemini Code Assist
Deepest hooks into existing enterprise toolchains — repo-wide search, PR review, and org-level policy that standalone editors don't match.
A library of agent loops, ready to steal
Worth bookmarking alongside the orchestration pattern in Section 04: Forward Future's Loop Library (signals.forwardfuture.com/loop-library) collects working agent-loop patterns — plan-act-observe variants, review-and-delegate splits like the Fable/Codex pairing above, and multi-agent debate patterns — as copy-paste starting points rather than theory. For teams building Hermes MoA presets or Claude Code routing blocks, it's a faster starting point than writing a control-flow pattern from scratch.
- Qwen3-Coder / Qwen3-Coder-Next — Alibaba's coding-specialized line, strong on Arabic-context codebases and self-hostable for data-residency requirements.
- Kimi K2.7 Code — Moonshot's coding-tuned release, competitive on agentic benchmarks at open-weight pricing.
- Devstral 2 — Mistral's dedicated coding model, positioned for self-hosted enterprise deployment.
For Conneqt's tooling stack
Given the Cursor/SpaceX/Grok 4.5 entanglement (Section 02) and Anthropic's rolling Fable 5 terms, the resilient stance this quarter is model-agnostic tooling — Cursor, Cline, or Aider — with Hermes MoA presets handling the multi-model routing underneath, rather than betting the whole workflow on one lab-native IDE.
07 · BY THE NUMBERS
The week, in numbers.
5
Days: Grok 4.5 to 2nd Fable Extension
2×
Fable 5 Access Extended This Week
$2/$6
Grok 4.5 Price (per M tokens)
~50%
Token Savings · Fable+Codex Loop
12
Free LLM Courses Curated
20–110
Credits per PowerPoint Task
08 · BOTTOM LINE
What to do this week.
FOR CODING WORKFLOWS
Try the Fable + Codex orchestration loop
Pick one genuinely hard, multi-step task this week — not a quick edit. Wire the CLAUDE.md routing block, run it against your normal single-model baseline, and compare token cost and output quality directly.
FOR CLAUDE PLANS
Don't build durable workflows on the Fable 5 promo
Budget for the $10/$50 credit-metered rate as the real cost of Fable 5, not the current 50%-included terms. Treat any extension as a bonus, not a plan.
FOR CLIENT DECKS
Pilot ChatGPT for PowerPoint on one pitch
Use it for first-draft assembly from discovery notes, then run your normal brand and Arabic-register review pass by hand. It's free for Business/Enterprise through August 6.
FOR TEAM TRAINING
Assign the LLM course track to new hires
Hugging Face LLM Course → AI Agents Course → LangChain Academy's LangGraph track. Free, self-paced, and maps directly onto the Hermes/MoA patterns Conneqt already runs.
FOR TOOLING RESILIENCE
Keep tooling model-agnostic
Given Cursor's SpaceX ownership and Anthropic's rolling Fable terms, route through Cline, Aider, or Hermes MoA presets rather than locking a whole workflow into one lab-native IDE.
FOR AGENT PATTERNS
Raid the Loop Library before writing from scratch
Check Forward Future's Loop Library for a working plan-act-observe or review-and-delegate pattern before building a new Hermes MoA preset or Claude Code routing block from zero.
The pattern this round
Five days, three labs, one visible flinch. Grok 4.5 undercut the frontier on price using data harvested from the coding tool its own parent company just bought. GPT-5.6 answered by merging two products into one and betting that "agent, not chatbot" is now the default UX for knowledge work, not just code. And Anthropic — which spent the last round explaining why policy and capability get negotiated separately — just demonstrated that competitive pressure gets negotiated in public, in real time, on X, hours before a deadline. The lesson compounds from Round 1: don't build anything durable on this week's terms from any single lab. Build the routing layer instead, and let the models compete underneath it.
