Deep Research Preview 04 2026
Deep Research Preview 04 2026 is a Google model available in SteelEngine. It supports a 1.05M token context window. Pricing starts at $2/1M input tokens and $12/1M output tokens. Key capabilities include Tool choice, Deep research, Max output 66k.
Runs long, multi-step research tasks that search, read, and synthesize sources on their own. Reads up to 1.05M tokens of context in one call and writes up to 66k tokens.
INPUT / 1M
$2
CACHED INPUT / 1M
N/A
OUTPUT / 1M
$12
CONTEXT WINDOW
1.05M
MAX OUTPUT
66k
RELEASED
PRICING
What Deep Research Preview 04 2026 costs to run.
List prices in USD per 1M tokens. Cached input applies when a request reuses a prompt the provider has already cached.
01 / INPUT
$2
per 1M input tokens
02 / CACHED INPUT
N/A
per 1M cached input tokens
03 / OUTPUT
$12
per 1M output tokens
LONG CONTEXT
SUPPORTED CAPABILITIES
What Deep Research Preview 04 2026 supports.
Capabilities show up as model settings on SteelEngine workflow blocks and agents, so teams can tune each step without changing providers.
CAPABILITIES
Tool choice, Deep research, Max output 66k, Memory off, Image input
CONTEXT WINDOW
1.05M
MAX OUTPUT TOKENS
66k
MODEL ID
deep-research-preview-04-2026
PROVIDER
MORE MODELS
More from Google
Compare nearby options from the same provider before you pick a default for your workflow.
RECOMMENDED
Gemini 4 Argon
Higher-priced reasoning model for complex planning and multi-step agent work. A recommended default among SteelEngine’s Google models.
IN / 1M
$2
OUT / 1M
$10
CONTEXT
1.05M
RECOMMENDED
Gemini 3.8 Flash
Mid-priced, latency-focused reasoning model for interactive agent steps. A recommended default among SteelEngine’s Google models. Released Sep 2, 2026.
IN / 1M
$0.75
OUT / 1M
$3.75
CONTEXT
1.05M
RECOMMENDED
Gemini 3.5 Flash
Mid-priced, latency-focused reasoning model for interactive agent steps. A recommended default among SteelEngine’s Google models. Released May 19, 2026.
IN / 1M
$1.5
OUT / 1M
$9
CONTEXT
1.05M
BUILD WITH STEELENGINE
Put Deep Research Preview 04 2026 to work.
Use this model in deterministic workflow steps or production agents, with approvals, guardrails, and a full audit trail around every call.