​

Z.ai, MiniMax, and Meta as Built-In Providers

• Self-hosted deployments can gate them with ZAI, MINIMAX, and META in PHOENIX_ALLOWED_PROVIDERS.
• New models on existing providers: gemini-3.8-flash on Google (20.6.0+) and gpt-6-astra on OpenAI (20.8.0+), with reasoning controls and pricing.

​

Filter Traces by Error Status and Latency

• error cuts both ways. true returns traces with at least one ERROR span, matching the UI indicator. false returns only clean traces, which is what you want when sampling healthy behavior for a dataset.
• Latency bounds are inclusive and apply to the trace, not its spans.
• The SDKs fail fast. An older server reports an upgrade requirement instead of silently returning unfiltered traces, and inverted bounds raise client-side.

from phoenix.client import Client

client = Client()

slow_failures = client.traces.get_traces(
    project_identifier="my-project",
    error=True,
    min_latency_ms=1000,
    limit=50,
)

import { getTraces } from "@arizeai/phoenix-client/traces";

const slowFailures = await getTraces({
  project: { projectName: "my-project" },
  error: true,
  minLatencyMs: 1000,
  limit: 50,
});

​

PXI: Tabbed Settings and Slash Command Suggestions

• Assistant settings are five linkable tabs — General, Tools, Permissions, Tracing & privacy, and Chats & data — each with its own URL under /settings/agents. Settings for capabilities your deployment has turned off stay hidden.
• The chat input suggests a slash command that fits what you have in context. Open a chat with a span attached and the placeholder points at the command for that situation.

​

MCP Server on FastMCP 4

• Progressive disclosure is gone. list_tool_groups and enable_tool_group no longer exist. With PHOENIX_ENABLE_MCP_CODE_MODE=false, every /v1 operation is advertised as its own tool.
• Analytics SQL errors read properly, so a bad query comes back as an actionable message.

​

Delete Traces from a Project over REST

• The time window is required. start_time and end_time bound the deletion as [start_time, end_time). Naive datetimes are read as UTC.
• Spans and orphaned sessions go with them. Sessions that still have traces outside the window survive.

curl -X DELETE \
  "$PHOENIX_ENDPOINT/v1/projects/my-project/traces?start_time=2026-08-01T00:00:00Z&end_time=2026-09-01T00:00:00Z" \
  -H "Authorization: Bearer $PHOENIX_API_KEY"

​

Additional Improvements

• Project token totals stop double-counting. Only leaf LLM spans count now, so frameworks that nest an LLM span inside another no longer inflate the dashboard headline.
• Metadata cells filter the table you are on. Clicking a metadata value in the traces table appends to the traces filter, not the span filter.
• Faster filtered trace and session queries. Filter lowering correlates every quantifier, which on PostgreSQL turns a statement timeout into an index probe.
• Annotation uploads validate every ID column up front, so an invalid document_position fails with a clear error instead of a confusing one later.
• PXI stops retrying doomed GitHub calls. A 401 or 403 is terminal and reports that the token’s scope or repository access needs checking.
• Experiments no longer finish early. A dispatched work item counts as in-flight from the moment it leaves the queue.
• Tool definitions carry their names on PXI’s own LLM spans, so its traces read cleanly in span details.
• Question counts hold steady while a PXI question streams in.