SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

GPT-Image-2.6 Adds SVG and GIF Output, but Skips the Batch API

0
·13 views

OpenAI released GPT-Image-2.6 Sunburst and Flare on November 4, adding SVG and animated GIF generation along with a new complexity setting. The Flare model supports only image generation and editing endpoints, with the Batch API marked unsupported, unlike the 2.5 Flare model, which offers a half-price batch option. Per-token pricing for the 2.6 models matches the 2.5 versions, though the newer models do not offer a dated snapshot for version pinning. A developer's forum comparison suggested medium-complexity images on 2.6 may use considerably more output tokens than on 2.5, though this is an unofficial estimate. Developers are advised to measure costs before switching model IDs in production.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

SentioGate Finishes Autonomous Coding Tasks 5x Faster Than OpenRouter Auto-Router in Benchmark

A benchmark comparing SentioGate and OpenRouter's auto-router on autonomous agent workflows found both completed all six assigned tasks, including test fixes, shell scripting, and debugging. SentioGate averaged 22.74 seconds per task, about 5.2 times faster than OpenRouter's 119.64 seconds, while OpenRouter's runs sometimes stalled for over two minutes. The authors attribute the gap to routing decisions based on real-time performance indices rather than list pricing or heuristics, which reduced wasted turns on conversational replies and retries. SentioGate also used about 30 percent fewer tokens on average, which matters because input context is billed on every turn of a multi-step run. The results come from the authors' own evaluation, and the methodology is described as reproducible with standard OpenAI client libraries.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Bluesky facet offsets count bytes, not characters, tripping up script posts

A developer's report of incorrect hashtag positions in a Bluesky API response turned out to be correct, because facet ranges are measured in UTF-8 bytes rather than characters. An emoji occupies one character but four bytes, which shifts every later offset. The same mismatch surfaced in September 2024, when a consumer found mention highlighting broken on posts with accented characters. The author shows that the offsets are accurate when read as byte positions, and that indexing by character produces the correct handles. The guidance is to compute facet ranges from the UTF-8 encoding of the text.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Migrating Bitnami Images: Missed Exporter Update and Adding pgvector to the Original Image

When an image source moves, every container in a Helm chart needs checking, not just the main one. The author switched the PostgreSQL and Redis images to the bitnamilegacy repository, but missed the metrics exporters, which kept pointing to the old source. The fix required separate values entries for each chart, and rendering confirmed the change. To add pgvector, the author built a multi-stage Dockerfile on the same Bitnami base image to keep changes small. Verification in an isolated environment confirmed existing data was preserved and vector operations worked, though the Legacy images are only a transitional option and are no longer updated.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Deterministic State Graphs with LangGraph and Redis for Reliable Banking AI

A technical article argues that linear LLM agent chains, modeled as acyclic pipelines, tend to fail in production banking environments. Errors such as API rate limits or database locks can cause hallucinated retries, unbounded loops, and heavy token costs, with no recovery path for lost state. The author proposes cyclical state machines built on LangGraph, with state checkpointed to Redis between each step so failed workflows can be rehydrated quickly. To address the memory bloat of JSON-based persistence, the design serializes checkpoints with MessagePack, and it adds grammar-based output validation to constrain model responses. The piece presents this architecture and its implementation code as a way to make regulated financial AI workflows more reliable.

0 comments Read more at DEV Community