Two PHP bugs found with AI are under attack, and DeepSeek's new Flash keeps its $0.15 price

The week on the beat

1. DeepSeek put V4.1 Flash 1017 behind its Flash API at the same price, and opened its search engine to agents

Calls to deepseek-flash now reach the new model with no code change, and DeepSeek's pricing page still lists $0.15/$0.60 per million off-peak, double that during weekday peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC). The weights are on Hugging Face under MIT, and DeepSeek reports shorter outputs than September's release. The new Search and Fetch APIs cost $0.75 per 1,000 successful requests each, and calls made through Claude Code, Codex or DeepSeek Harness are free until billing starts on November 1. MiniMax's M3.1 Flash, out on October 19, lists the same $0.15/$0.60 on its own pricing page. The same week, the US barred civilian federal agencies from using DeepSeek's products and services, so if you sell to them, read the order's scope before you ship anything built on it.

Read the story →

2. Meta released Muse Spark Open, an 816B open-weight model under its own license

It has 42B active parameters and a 1M context, and it takes text, images, video and audio and answers in text. Meta says it runs on two M5 Ultra Mac Studios with MLX; on GPUs the NVFP4 version needs about 700GB of memory. The Muse Model License requires a "Built with Muse Spark Open" notice, and if you sell access as a service with more than $20 million in yearly revenue you need a separate agreement with Meta. Hosted, it costs $1.25/$4.25 per million at Together, Fireworks and others, the same as Meta's list price for the closed Muse Spark 1.3. On Epoch's numbers it nearly matches 1.3 on the capabilities index (155.1 against 156.8) but sits well behind on the APEX agent benchmark (33.4% against 57.8%).

Read the story →

3. Two critical PHP bugs found with AI were exploited before the patch. Update now

PHP shipped 8.5.12, 8.4.27, 8.3.36 and 8.2.35 on October 21 to fix CVE-2026-109618 and CVE-2026-109619, both unauthenticated remote code execution rated 9.2, and the advisory says updating is the only complete fix. PHP 8.1 and older get no official release, only minimal backports as source branches. OpenAI's GPT-5.6 Sol found the critical flaw through its Patch the Planet program, and security firm Wiz found the same one independently with Gemini 4 Argon. Amazon Web Services logged exploit attempts days before the fix, and Wired reports the bugs are now broadly exploited. If anything you run sits on PHP, including WordPress, patch it today.

Read the story →

4. OpenAI's GPT-6 models are coming to Google Cloud, in private preview for now

GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna launch as a private preview for select enterprise customers, with sign-ups open, and usage runs through your Google Cloud bill and counts toward existing contracts. They will also appear in Gemini for Business and Google Antigravity, with no date given. That makes Google Cloud the first major cloud to sell Gemini, Claude and OpenAI's models in one place, so committed Google spend can now go to GPT-6.

Read the story →

5. Mistral released Voxtral 3: three transcription models, a text-to-speech model and a voice agent SDK

The SDK is built on LiveKit, covers more than 50 languages and lets an agent call your tools in the middle of a conversation. If you are stitching together separate speech-to-text, LLM and text-to-speech vendors for a voice agent, this puts the whole loop with one vendor.

Read the story →

6. OpenAI says two of its coding agents switched off the monitor meant to flag them, and GPT-6 Astra then denied it

In one case Astra disabled its safety monitor while copying reference code that had no monitor configured, and gave a false account when asked about it. OpenAI calls it a recurring pattern and says it does not yet have a satisfying explanation. It comes a week after Anthropic's report of Claude agents going off-script on real websites.

Read the story →

Model moves

• New: Beam 1.0 501B A23B (Reflection). The open model Reflection promised for October, with weights under Apache 2.0. It has 501B total and 23B active parameters and a 1M context, and it is built for coding and long agent runs. The cheapest host charges $0.25/$1 per 1M (Nebius, FP8). The benchmark claims so far are Reflection's own. Model page →
• New: Nemotron 3.5 Super VL (NVIDIA). An open-weight vision-language model with 120B total and 12B active parameters and a 1M context. The cheapest host charges $0.15/$0.50 per 1M (CoreWeave). Model page →
• New: Mercury 3 (Inception). A diffusion language model with a 1M context. List price $0.20/$0.70 per 1M, half that ($0.10/$0.35) during a launch discount with no end date given. Inception claims 1,068 tokens per second and wins over GPT-6 Luna and Claude Haiku 5.5 on 8 of its own 12 benchmarks; there are no independent numbers yet. Model page →
• New: DeepSeek V4.1 Flash 1017 (DeepSeek). $0.15/$0.60 per 1M off-peak and $0.30/$1.20 at weekday peak, with a 1M context and MIT weights. Model page →
• New: Muse Spark Open (Meta). Hosted at $1.25/$4.25 per 1M with a 1M context, under the Muse Model License. Model page →
• Sale ended: Mistral Large 4. The launch sale we reported on October 8 ($0.68/$2.09 per 1M) is over: Mistral's API now charges the $1.36/$4.18 list price, and its EU endpoint charges $1.496/$4.598. If you moved traffic onto it during the sale, that traffic now costs twice as much. Model page →
• Benchmark revision on Epoch's APEX listings, October 20. Muse Spark 1.2's reported score rose from 36.4% to 45.4%, and Kimi K3's fell from 50.6% to 47.8%. Model page →
• Restored Epoch scores. Our tracker missed about a month of Epoch updates after Epoch reorganized its data files. The fix on October 16 brought in three changes: DeepSeek-V3.2's APEX score went from 7.0% to 21.3%, Gemini 3.1 Pro's from 33.5% to 35.3%, and GPT-5.5's AIME 2024/2025 score from 84.4% to 100%. The 100% comes from Epoch's run of a pre-release GPT-5.5 at its highest effort setting. Model page →

Personal take

Two stories from this week go hand in hand. AI models discovered PHP bugs affecting Wikipedia and 4chan, and attackers exploited them before a patch was released. In the same week, OpenAI reported that two of its coding agents turned off the monitor designed to flag them, and one even denied it when questioned. Last week, Anthropic's agents submitted a fake police tip. It's good that these models are finding bugs, but I would still keep any agent away from files that control its own monitoring and permissions. If you use coding agents, check today to make sure they can't write to those files.

Until next Thursday, Anmol