# Grok 4.5 Breaks Into the Frontier as OpenAI Ships GPT-Live

*By AI High Signal Digest • July 9, 2026*

xAI's Grok 4.5 posted strong frontier agent benchmarks and aggressive pricing, while OpenAI rolled out GPT-Live and retracted its support for SWE-Bench Pro as a leading coding benchmark. Also inside: Cognition's SWE-1.7, major developer launches, and new infrastructure funding moves.

## Top Stories

*Why it matters: today's biggest shifts were in frontier model price-performance, voice interfaces, and trust in coding benchmarks.*

- **Grok 4.5 arrived with frontier-level agent results.** xAI says it is its first model trained specifically for coding and agents with Cursor [^1]. Artificial Analysis ranks it #4 on the Intelligence Index and says it is on par with GPT-5.5 on the Coding Agent Index at much lower cost; pricing is $2/$6 per 1M input/output tokens with a 500k context window [^2]. It also took the #1 spot on AutomationBench-AA at 51%, ahead of Claude Fable 5 and Claude Opus 4.8, at roughly a quarter of their cost per task [^3].

- **OpenAI rolled out GPT-Live in ChatGPT.** OpenAI introduced GPT-Live as a new generation of voice models; the system is full-duplex, so it can listen and speak at the same time, and it can delegate harder work to a frontier model in the background while keeping the conversation going [^4][^5][^6]. It is now fully rolled out to Go, Plus, and Pro users, with free rollout in progress and API access coming soon [^7][^8].

- **OpenAI also challenged a benchmark many labs use to market coding models.** After auditing SWE-Bench Pro, the company said 30% of tasks are broken, the eval is saturated at roughly a 70% noise ceiling, and it is retracting its prior recommendation to use the benchmark as a leading coding eval [^9][^10].

## Research & Innovation

*Why it matters: the most useful technical progress today was about cheaper coding performance, better model auditing, and more realistic views of agent failure.*

- **Cognition's SWE-1.7 pushed the cost/performance curve forward for coding.** Cognition says the model is within a few points of the strongest frontier models at a fraction of the cost and runs at 1000 tok/s [^11]. Built on RL pipeline improvements over a Kimi K2.7 base, it scores 42.3% on FrontierCode at $1.97 per task [^12].

- **D2D proposed a compact way to surface hidden bias.** The method is designed to find biases in fine-tuned models even when the auditor does not know the target topic, by distilling the difference from a base model into a 4M-parameter "Cartridge" that preserves the most coherent signal [^13].

- **A new Oxford-led survey mapped persistent agent failure modes.** Synthesizing 27 papers across 19 benchmarks, it groups failures into six clusters, including tool-use errors, planning failures, long-horizon degradation, coordination breakdowns, safety failures, and measurement problems, and argues that failures compound nonlinearly with task length [^14].

## Products & Launches

*Why it matters: launches are increasingly about turning models into usable workflows for developers, media teams, and edge applications.*

- **Runway Dev** launched as a new AI media platform for professional developers and enterprise teams, bundling zero-day model releases, pre-built endpoints, workflows as APIs, and real-time characters [^15][^16].

- **Moondream 3.1** landed on Cloudflare Workers AI, bringing image querying, captions, and object coordinates to edge inference with sub-second request times, including network round-trip time [^17][^18].

- **VS Code and GitHub Copilot** shipped a new set of agent features, including browser agent tools for web app validation, an Agents window for parallel workflows, bring-your-own-key support, and better cost/model visibility [^19][^20].

## Industry Moves

*Why it matters: capital and infrastructure continue to move toward open-model tooling and agent-native cloud stacks.*

- **Prime Intellect raised a $130M Series A** to build its Open Superintelligence Stack, led by Radical Ventures with NVIDIA, Intel Capital, and Dell Capital; the company says the stack lets users train, deploy, and continuously improve their own models [^21].

- **Modal raised a $355M Series C** around an agent-native cloud pitch built on sandboxes, elastic inference, GPU snapshotting, and support for up to 100,000 RL rollouts [^22][^23].

- **Together Compute introduced Provisioned Throughput** for frontier open models, offering reserved inference capacity, token-based pricing, a 99% uptime SLA, and initial support for MiniMax M3 and GLM-5.2 [^24].

## Policy & Regulation

*Why it matters: concrete government actions are now shaping who gets access to large-scale training capacity.*

- **China approved purchases of 200,000 NVIDIA H200 chips** by DeepSeek, ByteDance, Alibaba, and others for training [^25]. A separate estimate put the implied spend at about $5B using $25,000 per GPU [^26].

## Quick Takes

*Why it matters: a few smaller updates still sharpened the picture on model quality, open-weight progress, and multimodal tooling.*

- GPT-5.6 Sol, Terra, and Luna were added to the Codex codebase, where Sol is described as OpenAI's "most capable model yet" [^27].
- Google's Nano Banana 2 Lite debuted at #5 on Artificial Analysis's text-to-image leaderboard, generates 1K images in about 3.4 seconds, and costs half as much as Nano Banana 2 [^28].
- Z.ai's GLM-5.2 scored 152 on the Epoch Capabilities Index, the highest among open-weight models Epoch has evaluated [^29].
- MOSS-Transcribe-Diarize-0.9B open-sourced a single-pass model for up to ~90 minutes of multi-speaker transcription with timestamps and speaker labels [^30][^31].

---

### Sources

[^1]: [𝕏 post by @SpaceXAI](https://x.com/SpaceXAI/status/2074915721684086811)
[^2]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2074956932289282087)
[^3]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2075047187525034114)
[^4]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2074907025537224840)
[^5]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2074907028733178359)
[^6]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2074907033577693636)
[^7]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2075019750569378007)
[^8]: [𝕏 post by @juberti](https://x.com/juberti/status/2074908711773212895)
[^9]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2074972179385720836)
[^10]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2074958149426241894)
[^11]: [𝕏 post by @cognition](https://x.com/cognition/status/2074882968770728416)
[^12]: [𝕏 post by @cognition](https://x.com/cognition/status/2074882970935034309)
[^13]: [𝕏 post by @Azaliamirh](https://x.com/Azaliamirh/status/2074942191982559516)
[^14]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2074874153245814864)
[^15]: [𝕏 post by @runwayml](https://x.com/runwayml/status/2074855836417974708)
[^16]: [𝕏 post by @c_valenzuelab](https://x.com/c_valenzuelab/status/2074869306521698552)
[^17]: [𝕏 post by @CloudflareDev](https://x.com/CloudflareDev/status/2074964416227877258)
[^18]: [𝕏 post by @vikhyatk](https://x.com/vikhyatk/status/2074971060110545315)
[^19]: [𝕏 post by @code](https://x.com/code/status/2074943855967821851)
[^20]: [𝕏 post by @pierceboggan](https://x.com/pierceboggan/status/2074908789716009285)
[^21]: [𝕏 post by @PrimeIntellect](https://x.com/PrimeIntellect/status/2074899489190785419)
[^22]: [𝕏 post by @latentspacepod](https://x.com/latentspacepod/status/2075007025600504128)
[^23]: [𝕏 post by @akshat_b](https://x.com/akshat_b/status/2075036759571022298)
[^24]: [𝕏 post by @togethercompute](https://x.com/togethercompute/status/2074756085785722930)
[^25]: [𝕏 post by @jukan05](https://x.com/jukan05/status/2074871701616329193)
[^26]: [𝕏 post by @teortaxesTex](https://x.com/teortaxesTex/status/2074872962734301339)
[^27]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2075046739682168977)
[^28]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2074917317591564619)
[^29]: [𝕏 post by @EpochAIResearch](https://x.com/EpochAIResearch/status/2074894535558300103)
[^30]: [𝕏 post by @MosiAI_Official](https://x.com/MosiAI_Official/status/2075059157443756245)
[^31]: [𝕏 post by @vllm_project](https://x.com/vllm_project/status/2075063247837757474)