# Mythos Safeguards, DeepSeek Price Cuts, and Google’s Agent Expansion

*By AI High Signal Digest • May 23, 2026*

Anthropic linked stronger cyber capabilities to tighter safeguards, DeepSeek pushed frontier pricing lower with a permanent V4 Pro cut, and Google expanded into persistent agents and conversational video. Also inside: efficient agent research, notable product launches in speech and security, and two meaningful policy moves.

## Top Stories

*Why it matters: today’s biggest signals were tighter safety limits on frontier cyber models, faster price compression at the frontier, and broader consumer AI productization.*

- **Anthropic tied frontier cyber capability to tighter release controls.** The company said Project Glasswing and its partners have already found more than 10,000 high- or critical-severity vulnerabilities in essential software, and warned that the software industry will need to adapt to the volume that models like Claude Mythos Preview can find [^1][^2]. Anthropic also said it wants “far stronger safeguards” before a general release of Mythos-class models [^3]. In external benchmark posts, Mythos was reported to beat GPT-5.5 on SWE-bench Pro, HLE, UK AISI cyber ranges, and exploit benchmarks [^4].

- **DeepSeek made its V4 Pro price cut permanent.** First-party pricing is now $0.435 per 1M input tokens, $0.87 output, and $0.0036 cached input, with a blended price around $0.18 per 1M [^5]. Artificial Analysis said the move puts V4 Pro on the Intelligence Index vs. cost Pareto frontier, with index cost around $268 versus $892 for Gemini 3.1 Pro Preview, $3,357 for GPT-5.5, and $5,117 for Claude Opus 4.7 [^5].

- **Google expanded beyond chat into persistent agents and conversational video.** Gemini Spark is framed as a 24/7 personal AI agent for recurring tasks, new skills, and end-to-end workflows [^6]. Gemini Omni lets users create and edit video through natural language, with custom avatars, multimodal inputs, and physics-aware scene consistency [^7][^8][^9][^10][^11].

## Research & Innovation

*Why it matters: the most useful research updates were about lowering agent cost, improving training efficiency, and measuring model quality more honestly.*

- **Agent workflows may be getting compiled into weights.** A paper highlighted by DAIR says full agentic workflows can be distilled into a model at roughly 100x lower inference cost while preserving near-frontier task quality [^12].

- **Introspective X Training targets better performance per FLOP.** The method annotates data with model-generated critiques up front and reported up to 2.8x FLOP efficiency plus 5–10 point gains, especially in math and code, across training stages on 8B models [^13].

- **A new eval paper argues loss is a weak proxy for reasoning quality.** “Forecasting Downstream Performance of LLMs With Proxy Metrics” says cross-entropy poorly predicts downstream reasoning performance and proposes proxy metrics over expert reasoning traces instead [^14].

## Products & Launches

*Why it matters: new launches focused less on standalone chat and more on developer infrastructure, speech quality, and security tooling.*

- **Google’s Managed Agents + Interactions API** gives an agent a secure hosted Linux sandbox for code execution and memory management through a single API flow [^15].

- **Cartesia’s Sonic-3.5** took the top spot on the Artificial Analysis Speech Arena with Elo 1,218, ahead of Inworld Realtime TTS 1.5 Max and Gemini 3.1 Flash TTS; it supports 42 languages and 500+ voices [^16].

- **Perplexity open-sourced Bumblebee**, a read-only scanner for macOS and Linux that checks developer machines for risky packages, extensions, and AI tool configs, and can trigger deeper scans when new supply-chain risks emerge [^17][^18].

## Industry Moves

*Why it matters: capital and partnerships continue to flow toward labs and deployments that can turn model capability into durable distribution.*

- **DeepSeek’s financing round is advancing alongside an open-source stance.** Posts citing Bloomberg said the company is moving ahead with a roughly $10B financing round, while founder Liang Wenfeng told investors DeepSeek has no interest in short-term commercialization and will remain open-source [^19][^20].

- **Google DeepMind expanded its partnership with Singapore** to help deploy AI at scale in scientific discovery, pandemic preparedness, and healthcare [^21].

- **AI media production is moving into core operating budgets.** One major e-commerce company said it now generates 75% of its visual media with AI using Runway, reproducing projects that once cost $800K for under $10K and shifting a $5–6M annual production budget toward AI workflows [^22].

## Policy & Regulation

*Why it matters: governments and regulators are starting to shape who can own AI assets and what AI marketing claims can survive scrutiny.*

- **China halted Meta’s planned acquisition of Manus**, signaling tighter state control over strategically important AI technology and disrupting the common startup strategy of relocating abroad for Western capital and partnerships [^23].

- **The FTC settled active-listening AI marketing claims.** Cox Media Group and two other firms will pay nearly $1 million over allegations they deceived customers about an AI-powered ad-targeting service [^24].

## Quick Takes

*Why it matters: a few smaller updates sharpened the picture on provenance, search efficiency, world models, and robotics.*

- **SynthID** is expanding to more partners, and Google added new AI-content detection paths in Gemini App and Search [^25].
- **MiniMax Agent** switched its default search from Serper to Perplexity, cutting tool calls 45%, token usage 42%, and total cost 27% while raising pass rate 2% [^26].
- **Project Genie** can now turn Google Maps Street View locations into interactive worlds for eligible Google AI Ultra subscribers [^27][^28].
- **Figure** said a recent logistics deployment ran continuously for 200 hours without failure [^29].

---

### Sources

[^1]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2057909102542549503)
[^2]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2057909104090169464)
[^3]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2057911231076237661)
[^4]: [𝕏 post by @scaling01](https://x.com/scaling01/status/2057980199707214323)
[^5]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2058021452465799403)
[^6]: [𝕏 post by @Google](https://x.com/Google/status/2057841803550683336)
[^7]: [𝕏 post by @Google](https://x.com/Google/status/2057881884219035752)
[^8]: [𝕏 post by @Google](https://x.com/Google/status/2057881895241744749)
[^9]: [𝕏 post by @Google](https://x.com/Google/status/2057881888702845440)
[^10]: [𝕏 post by @Google](https://x.com/Google/status/2057881890544037956)
[^11]: [𝕏 post by @Google](https://x.com/Google/status/2057881886345805927)
[^12]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2057846601843146760)
[^13]: [𝕏 post by @rajammanabrolu](https://x.com/rajammanabrolu/status/2057893818415157605)
[^14]: [𝕏 post by @arkil_patel](https://x.com/arkil_patel/status/2057819898349572156)
[^15]: [𝕏 post by @_philschmid](https://x.com/_philschmid/status/2057833963633418426)
[^16]: [𝕏 post by @ArtificialAnlys](https://x.com/ArtificialAnlys/status/2057878247782908109)
[^17]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2057869990536360334)
[^18]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2057870003060498612)
[^19]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2057907839839154302)
[^20]: [𝕏 post by @luluyilun](https://x.com/luluyilun/status/2057664819176857853)
[^21]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2057985225100235022)
[^22]: [𝕏 post by @c_valenzuelab](https://x.com/c_valenzuelab/status/2058024258115346733)
[^23]: [𝕏 post by @DeepLearningAI](https://x.com/DeepLearningAI/status/2057959692907737397)
[^24]: [𝕏 post by @FTC](https://x.com/FTC/status/2057517453215666267)
[^25]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2057898089621459434)
[^26]: [𝕏 post by @MiniMaxAgent](https://x.com/MiniMaxAgent/status/2057844100205031779)
[^27]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2057842131142590512)
[^28]: [𝕏 post by @GoogleDeepMind](https://x.com/GoogleDeepMind/status/2057842133046903086)
[^29]: [𝕏 post by @adcock_brett](https://x.com/adcock_brett/status/2057907052140486938)