Skip to main content
Back

OpenAI Unveils GPT-5: Breaking Benchmarks + 256K Context

GPT-5 with 256k context, new DeepMind and Anthropic models, ElevenLabs music gen, and more..

August 7, 2025 4 min read

Hey

Welcome to AlphaSignal, the most read source of news by AI engineers and researchers. Every day, we identity and summarize the top 1% of news, papers, models, and repos, so you’re always up to date.


Here’s today’s roundup:

./SUMMARY

Read time: 4 min 19 sec

Top News

▪ OpenAI Unveils GPT-5: New Benchmark for Code, 256K Context

ASAPP

▪ New customer service AI agent that approve responses, monitor chats, prevent ticket loss.

Trending Signals

▪ DeepMind’s new Genie 3 can create 3D environments with real-time interactivity

▪ Anthropic ships Claude Opus 4, beats o3 with improved coding performance

▪ ElevenLabs rolls out full-track music generation tool with sound and lyric control

▪ Google introduces Guided Learning to Gemini for step-by-step deep dives

▪ Anthropic adds automated security reviews to Claude Code workflows

AssemblyAI

▪ Speed up audio workflows with AssemblyAI API

Top Papers

▪ OpenAI evaluates risk of malicious fine-tuning in GPT-OSS

▪ How AI agents could handle complex tasks for you on the web

▪ Google proposes diffusion-based agent for better AI report generation

How To

▪ Fine-tune GPT-OSS-20B on reasoning tasks in six simple steps by OpenAI

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

./TOP_NEWS

OpenAI Unveils GPT-5: New Benchmark for Code, Context, and Automation

22,000 Likes

OpenAI finally released the highly anticipated GPT-5, introducing its most advanced language model to date for ChatGPT and API users.


The launch upgrades the core engine powering millions of AI workflows, raising benchmarks in coding, automation, and large-context tasks. “GPT-5 is a significant step along the path to AGI… a model that is generally intelligent,” said CEO Sam Altman.


With a 256,000-token context window, smarter model routing, and major improvements in code and science performance, GPT-5 sets a new bar for developers and researchers.


What GPT-5 Excels At
GPT-5 outperforms all previous OpenAI models in key technical areas:

  • Coding: Achieves 74.9% on SWE-bench Verified, above Claude Opus 4.1 (74.5%) and Gemini 2.5 Pro (59.6%).

  • Multilingual code editing: Reaches 88% on Aider Polyglot.

  • PhD-level science: Scores 89.4% on GPQA Diamond (GPT-5 Pro variant).

  • Long-context tasks: Handles up to 256,000 tokens, enabling processing of full codebases and large documents.

  • Agentic automation: Manages complex, multi-step tasks with tool use and instruction following.

Key Improvements Over Previous Models
OpenAI increases GPT-5’s capabilities in several core areas:

  • Expands the context window well beyond previous limits.

  • Automates model selection for each query, allocating more compute to hard problems.

  • Improves code reliability and error handling, reducing developer friction.

  • Offers deeper reasoning control in the Pro tier with “GPT-5 Thinking.”

  • Adds user controls for verbosity and output length via API.

What Remains Good, Not Best-in-Class
GPT-5 delivers strong but not category-leading results in some areas:

  • Some agentic benchmarks show parity or small gains versus top competitors.

  • Everyday writing quality improves modestly, but not dramatically.

  • Output remains reliable, but full validation is still necessary for critical use cases.

What Did Not Improve or Is Still Lacking
Some features remain unchanged or unavailable:

  • No public fine-tuning at launch.

  • Model does not update after deployment (no continuous learning).

  • Automated routing can misjudge query complexity without explicit user prompt.

Pricing and Migration

  • gpt-5: $1.25 per million input tokens, $10 per million output tokens.

  • gpt-5-mini: $0.25 and $2.

  • gpt-5-nano: $0.05 and $0.40.

  • Pro tier: $200/month with unlimited usage and reasoning depth.

  • Migration: Update model name in API; test workflows for changes in output

Read more

Accelerate Reliable Customer Service AI Deployment With Real-Time Monitoring and QA

ASAPP gives you the tools for full control over your AI customer service agent.


Get safety and security by design, real-time expert input, and test every scenario before going live.

  • Review and approve agent responses instantly while training AI faster

  • Simulate interactions in a test environment to ensure safety and reliability

  • Monitor all AI conversation activity for transparency and QA at scale

  • Deploy the most capable generative AI agent to resolve issues faster

Try now

partner with us

./TOP_SIGNALS

DeepMind launches Genie 3, a foundation world model for real-time simulation and embodied agent training

46,385 Likes

Anthropic presents Opus 4.1, improving real-world coding, data analysis, and multi-file refactoring

12,495 Likes

ElevenLabs announces Eleven Music, which generates full-length songs with editable lyrics, 44.1 kHz audio, and multilingual vocals

11,385 Likes

Google rolls out Guided Learning to Gemini with step-by-step reasoning for structured topic explanation; offers free Pro Plan for students

8,395 Likes

Anthropic adds an AI-powered vulnerability detection tool for Claude Code using a command and GitHub integration

11,495 Likes

Speed up Speech + LLM App Development with AssemblyAI API

Transcribe audio, detect speakers, and automate language detection with AssemblyAI. Get structured outputs for summarization, sentiment, and QA workflows.

Use built-in diarization for conversation analytics.

Try the API for free ↗️

./TOP_PAPERS

OpenAI Evaluates Risk of Malicious Fine-Tuning in GPT-OSS

2,184 Likes

OpenAI tests whether GPT-OSS models can be fine-tuned for dangerous tasks like bioweapon design. Even under adversarial setups, the open weights models fail to reach high-risk thresholds, unlike proprietary frontier models.

How AI Agents could Handle Complex Tasks for You on the Web

1,048 Likes

The paper defines the Agentic Web, where AI agents execute complex tasks for users through planning and coordination. It explores key dimensions, intelligence, interaction, and economics. It identifies system challenges and outlines applications, societal risks, and governance needs.


Google Proposes Diffusion-Based Agent for Better AI Report Generation

1,194 Likes

The study addresses limitations in long‑form, multi‑hop report generation by research agents. It proposes a diffusion-based report generation approach that starts with a draft, refines it via retrieval and self‑evolution. It outperforms existing agents on key benchmarks.

./HOW_TO

Fine-Tune GPT-OSS-20B on Reasoning Tasks by OpenAI

You’ll train the model on multilingual chain-of-thought data using the Hugging Face ecosystem. Here is a gist:


Step 1
Install required libraries and log in to Hugging Face. 


from huggingface_hub import notebook_login
notebook_login()

Step 2
Load the dataset.


from datasets import load_dataset
dataset = load_dataset("HuggingFaceH4/Multilingual-Thinking",
                                      split="train")

Step 3
Load the openai/gpt-oss-20b model. Configure LoRA (low‑rank adaptation) to reduce memory usage.

Step 4

Run training on the multilingual reasoning dataset using Hugging Face's TRL (Transformer Reinforcement Learning library).

Step 5

Send prompts in one language, request a chain‑of‑thought in another. For example: ask in Spanish, reason in German, and get the final answer in Spanish.

Step 6

Save the fine‑tuned model. Push it to the Hugging Face Hub for reuse.

Jupyter notebook

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us